fix

* feat: accept pdf, images and office files

* refactor: rename FileConverter to FileParser, simplify file retain API

- Rename engine/converters/ → engine/parsers/, FileConverter → FileParser,
  ConverterRegistry → FileParserRegistry, MarkitdownConverter → MarkitdownParser
- Rename env var HINDSIGHT_API_FILE_CONVERTER → HINDSIGHT_API_FILE_PARSER
- Remove async/document_tags params from FileRetainRequest (always async now)
- Add retain_files() to Python Hindsight client and retainFiles() to TypeScript client
- Add sample.pdf to doc examples for working file upload demonstrations
- Update test_file_retain.py to use new parser names and always-async behavior
- Fix Go client missing os import in api_files.go
- Simplify postgresql.py storage to minimal schema

* fix: update rust CLI tests to use is_supported_file instead of is_text_file

* fix: patch Go api_files.go to add missing 'os' import after generation

* fix: insert 'os' import after 'net/url' in api_files.go patch for correct position

* chore: regenerate OpenAPI spec and clients (converter→parser description update)

.env.example

@@ -2,7 +2,7 @@
 # Copy this file to .env and fill in your values
 
 # LLM Configuration (Required)
-# Supported providers: openai, groq, ollama, gemini, anthropic, lmstudio
+# Supported providers: openai, groq, ollama, gemini, anthropic, lmstudio, vertexai
 HINDSIGHT_API_LLM_PROVIDER=openai
 HINDSIGHT_API_LLM_API_KEY=your-api-key-here
 HINDSIGHT_API_LLM_MODEL=o3-mini
@@ -13,6 +13,13 @@ HINDSIGHT_API_LLM_BASE_URL=https://api.openai.com/v1
 # HINDSIGHT_API_LLM_API_KEY=your-anthropic-api-key
 # HINDSIGHT_API_LLM_MODEL=claude-sonnet-4-20250514
 
+# Example: Google Vertex AI configuration
+# HINDSIGHT_API_LLM_PROVIDER=vertexai
+# HINDSIGHT_API_LLM_MODEL=google/gemini-2.0-flash-001
+# HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID=your-gcp-project-id
+# HINDSIGHT_API_LLM_VERTEXAI_REGION=us-central1
+# HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY=/path/to/service-account-key.json  # Optional, uses ADC if not set
+
 # Example: LM Studio local configuration (Qwen 2.5 32B recommended)
 # HINDSIGHT_API_LLM_PROVIDER=lmstudio
 # HINDSIGHT_API_LLM_API_KEY=lmstudio
@@ -24,8 +31,21 @@ HINDSIGHT_API_HOST=0.0.0.0
 HINDSIGHT_API_PORT=8888
 HINDSIGHT_API_LOG_LEVEL=info
 
+# Base Path / Reverse Proxy Support (Optional)
+# Set these when deploying behind a reverse proxy with path-based routing
+# Example: To deploy at example.com/hindsight/, set both to "/hindsight"
+# HINDSIGHT_API_BASE_PATH=/hindsight
+# NEXT_PUBLIC_BASE_PATH=/hindsight
+
 # Database (Optional - uses embedded pg0 by default)
 # HINDSIGHT_API_DATABASE_URL=postgresql://user:pass@host:5432/db
+# HINDSIGHT_API_DATABASE_SCHEMA=public  # PostgreSQL schema name (default: public)
+
+# Vector Extension (Optional - uses pgvector by default)
+# Options: "pgvector" (default), "vchord", "pgvectorscale" (DiskANN)
+# HINDSIGHT_API_VECTOR_EXTENSION=pgvector
+# For Azure PostgreSQL with DiskANN:
+# HINDSIGHT_API_VECTOR_EXTENSION=pgvectorscale  # Auto-detects pg_diskann on Azure
 
 # Embeddings Configuration (Optional - uses local by default)
 # Provider: "local" (default) or "tei" (HuggingFace Text Embeddings Inference)
@@ -42,3 +62,18 @@ HINDSIGHT_API_LOG_LEVEL=info
 # HINDSIGHT_API_RERANKER_LOCAL_MODEL=cross-encoder/ms-marco-MiniLM-L-6-v2
 # For TEI provider:
 # HINDSIGHT_API_RERANKER_TEI_URL=http://localhost:8081
+
+# Observability & Tracing (Optional - disabled by default)
+# Enable OpenTelemetry tracing for LLM calls (GenAI semantic conventions)
+# HINDSIGHT_API_OTEL_TRACES_ENABLED=true
+#
+# Local development with Grafana LGTM stack (recommended - see scripts/dev/grafana/README.md)
+# HINDSIGHT_API_OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4318
+#
+# Cloud backends (Grafana Cloud, Langfuse, DataDog, etc.)
+# HINDSIGHT_API_OTEL_EXPORTER_OTLP_ENDPOINT=https://your-backend-url
+# HINDSIGHT_API_OTEL_EXPORTER_OTLP_HEADERS="Authorization=Bearer your-token"
+#
+# Custom service name and environment (optional, defaults: hindsight-api, development)
+# HINDSIGHT_API_OTEL_SERVICE_NAME=hindsight-production
+# HINDSIGHT_API_OTEL_DEPLOYMENT_ENVIRONMENT=production

.github/workflows/release.yml

@@ -139,6 +139,104 @@ jobs:
         path: hindsight-clients/typescript/*.tgz
         retention-days: 1
 
+  release-openclaw-integration:
+    runs-on: ubuntu-latest
+    environment: npm
+
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Set up Node.js
+      uses: actions/setup-node@v4
+      with:
+        node-version: '22'
+        registry-url: 'https://registry.npmjs.org'
+
+    - name: Install dependencies
+      working-directory: ./hindsight-integrations/openclaw
+      run: npm ci
+
+    - name: Build
+      working-directory: ./hindsight-integrations/openclaw
+      run: npm run build
+
+    - name: Publish to npm
+      working-directory: ./hindsight-integrations/openclaw
+      run: |
+        set +e
+        OUTPUT=$(npm publish --access public 2>&1)
+        EXIT_CODE=$?
+        echo "$OUTPUT"
+        if [ $EXIT_CODE -ne 0 ]; then
+          if echo "$OUTPUT" | grep -q "cannot publish over"; then
+            echo "Package version already published, skipping..."
+            exit 0
+          fi
+          exit $EXIT_CODE
+        fi
+      env:
+        NODE_AUTH_TOKEN: ${{ secrets.NPM_TOKEN }}
+
+    - name: Pack for GitHub release
+      working-directory: ./hindsight-integrations/openclaw
+      run: npm pack
+
+    - name: Upload artifacts
+      uses: actions/upload-artifact@v4
+      with:
+        name: openclaw-integration
+        path: hindsight-integrations/openclaw/*.tgz
+        retention-days: 1
+
+  release-ai-sdk-integration:
+    runs-on: ubuntu-latest
+    environment: npm
+
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Set up Node.js
+      uses: actions/setup-node@v4
+      with:
+        node-version: '22'
+        registry-url: 'https://registry.npmjs.org'
+
+    - name: Install dependencies
+      working-directory: ./hindsight-integrations/ai-sdk
+      run: npm ci
+
+    - name: Build
+      working-directory: ./hindsight-integrations/ai-sdk
+      run: npm run build
+
+    - name: Publish to npm
+      working-directory: ./hindsight-integrations/ai-sdk
+      run: |
+        set +e
+        OUTPUT=$(npm publish --access public 2>&1)
+        EXIT_CODE=$?
+        echo "$OUTPUT"
+        if [ $EXIT_CODE -ne 0 ]; then
+          if echo "$OUTPUT" | grep -q "cannot publish over"; then
+            echo "Package version already published, skipping..."
+            exit 0
+          fi
+          exit $EXIT_CODE
+        fi
+      env:
+        NODE_AUTH_TOKEN: ${{ secrets.NPM_TOKEN }}
+
+    - name: Pack for GitHub release
+      working-directory: ./hindsight-integrations/ai-sdk
+      run: npm pack
+
+    - name: Upload artifacts
+      uses: actions/upload-artifact@v4
+      with:
+        name: ai-sdk-integration
+        path: hindsight-integrations/ai-sdk/*.tgz
+        retention-days: 1
+
   release-control-plane:
     runs-on: ubuntu-latest
     environment: npm
@@ -242,6 +340,7 @@ jobs:
         retention-days: 1
 
   release-docker-images:
+    name: Release Docker (${{ matrix.image_name }}${{ matrix.tag_suffix }})
     runs-on: ubuntu-latest
     permissions:
       contents: read
@@ -251,10 +350,28 @@ jobs:
         include:
           - target: api-only
             image_name: hindsight-api
+            tag_suffix: ""
+            build_args: ""
+          - target: api-only
+            image_name: hindsight-api
+            tag_suffix: "-slim"
+            build_args: |
+              INCLUDE_LOCAL_MODELS=false
+              PRELOAD_ML_MODELS=false
           - target: cp-only
             image_name: hindsight-control-plane
+            tag_suffix: ""
+            build_args: ""
           - target: standalone
             image_name: hindsight
+            tag_suffix: ""
+            build_args: ""
+          - target: standalone
+            image_name: hindsight
+            tag_suffix: "-slim"
+            build_args: |
+              INCLUDE_LOCAL_MODELS=false
+              PRELOAD_ML_MODELS=false
 
     steps:
     - uses: actions/checkout@v4
@@ -292,6 +409,9 @@ jobs:
       uses: docker/metadata-action@v5
       with:
         images: ghcr.io/${{ github.repository_owner }}/${{ matrix.image_name }}
+        flavor: |
+          latest=auto
+          suffix=${{ matrix.tag_suffix }}
         tags: |
           type=semver,pattern={{version}},value=${{ steps.get_version.outputs.VERSION }}
           type=semver,pattern={{major}}.{{minor}},value=${{ steps.get_version.outputs.VERSION }}
@@ -317,7 +437,7 @@ jobs:
     # - name: Smoke test - verify container starts
     #   env:
     #     GROQ_API_KEY: ${{ secrets.GROQ_API_KEY }}
-    #   run: ./scripts/docker-smoke-test.sh "${{ matrix.image_name }}:test" "${{ matrix.target }}"
+    #   run: ./docker/test-image.sh "${{ matrix.image_name }}:test" "${{ matrix.target }}"
 
     # Build multi-platform and push to release tags
     - name: Build and push release images
@@ -326,6 +446,7 @@ jobs:
         context: .
         file: docker/standalone/Dockerfile
         target: ${{ matrix.target }}
+        build-args: ${{ matrix.build_args }}
         push: true
         platforms: linux/amd64,linux/arm64
         tags: ${{ steps.meta.outputs.tags }}
@@ -366,7 +487,7 @@ jobs:
 
   create-github-release:
     runs-on: ubuntu-latest
-    needs: [release-python-packages, release-typescript-client, release-control-plane, release-rust-cli, release-docker-images, release-helm-chart]
+    needs: [release-python-packages, release-typescript-client, release-openclaw-integration, release-ai-sdk-integration, release-control-plane, release-rust-cli, release-docker-images, release-helm-chart]
     permissions:
       contents: write
 
@@ -389,6 +510,18 @@ jobs:
         name: typescript-client
         path: ./artifacts/typescript-client
 
+    - name: Download OpenClaw Integration
+      uses: actions/download-artifact@v4
+      with:
+        name: openclaw-integration
+        path: ./artifacts/openclaw-integration
+
+    - name: Download AI SDK Integration
+      uses: actions/download-artifact@v4
+      with:
+        name: ai-sdk-integration
+        path: ./artifacts/ai-sdk-integration
+
     - name: Download Control Plane
       uses: actions/download-artifact@v4
       with:
@@ -430,6 +563,10 @@ jobs:
         cp artifacts/python-packages/hindsight-embed/dist/* release-assets/ || true
         # TypeScript client
         cp artifacts/typescript-client/*.tgz release-assets/ || true
+        # OpenClaw Integration
+        cp artifacts/openclaw-integration/*.tgz release-assets/ || true
+        # AI SDK Integration
+        cp artifacts/ai-sdk-integration/*.tgz release-assets/ || true
         # Control Plane
         cp artifacts/control-plane/*.tgz release-assets/ || true
         # Rust CLI binaries

.github/workflows/test.yml

@@ -9,42 +9,11 @@ concurrency:
   cancel-in-progress: true
 
 jobs:
-  build-python-packages:
-    runs-on: ubuntu-latest
-    strategy:
-      matrix:
-        include:
-          - name: hindsight-all
-            path: hindsight
-          - name: hindsight-api
-            path: hindsight-api
-          - name: hindsight-client
-            path: hindsight-clients/python
-          - name: hindsight-embed
-            path: hindsight-embed
-
-    steps:
-    - uses: actions/checkout@v4
-
-    - name: Install uv
-      uses: astral-sh/setup-uv@v5
-      with:
-        enable-cache: true
-
-    - name: Set up Python
-      uses: actions/setup-python@v5
-      with:
-        python-version-file: ".python-version"
-
-    - name: Build ${{ matrix.name }}
-      working-directory: ./${{ matrix.path }}
-      run: uv build
-
   build-api-python-versions:
     runs-on: ubuntu-latest
     strategy:
       matrix:
-        python-version: ['3.11', '3.12', '3.13']
+        python-version: ['3.11', '3.12', '3.13', '3.14']
 
     steps:
     - uses: actions/checkout@v4
@@ -82,6 +51,52 @@ jobs:
     - name: Build TypeScript client
       run: npm run build --workspace=hindsight-clients/typescript
 
+  build-openclaw-integration:
+    runs-on: ubuntu-latest
+
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Set up Node.js
+      uses: actions/setup-node@v4
+      with:
+        node-version: '22'
+
+    - name: Install dependencies
+      working-directory: ./hindsight-integrations/openclaw
+      run: npm ci
+
+    - name: Run tests
+      working-directory: ./hindsight-integrations/openclaw
+      run: npm test
+
+    - name: Build
+      working-directory: ./hindsight-integrations/openclaw
+      run: npm run build
+
+  build-ai-sdk-integration:
+    runs-on: ubuntu-latest
+
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Set up Node.js
+      uses: actions/setup-node@v4
+      with:
+        node-version: '22'
+
+    - name: Install dependencies
+      working-directory: ./hindsight-integrations/ai-sdk
+      run: npm ci
+
+    - name: Run tests
+      working-directory: ./hindsight-integrations/ai-sdk
+      run: npm test
+
+    - name: Build
+      working-directory: ./hindsight-integrations/ai-sdk
+      run: npm run build
+
   build-control-plane:
     runs-on: ubuntu-latest
 
@@ -153,8 +168,15 @@ jobs:
     - name: Build docs
       run: npm run build --workspace=hindsight-docs
 
-  build-rust-cli:
+  test-rust-cli:
     runs-on: ubuntu-latest
+    env:
+      HINDSIGHT_API_LLM_PROVIDER: groq
+      HINDSIGHT_API_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
+      HINDSIGHT_API_LLM_MODEL: openai/gpt-oss-20b
+      HINDSIGHT_API_URL: http://localhost:8888
+      GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+      UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu
 
     steps:
     - uses: actions/checkout@v4
@@ -171,6 +193,10 @@ jobs:
           hindsight-cli/target
         key: ${{ runner.os }}-cargo-${{ hashFiles('**/Cargo.lock') }}
 
+    - name: Run unit tests
+      working-directory: hindsight-cli
+      run: cargo test
+
     - name: Build CLI
       working-directory: hindsight-cli
       run: cargo build --release
@@ -182,29 +208,6 @@ jobs:
         path: hindsight-cli/target/release/hindsight
         retention-days: 1
 
-  test-rust-cli:
-    runs-on: ubuntu-latest
-    needs: build-rust-cli
-    env:
-      HINDSIGHT_API_LLM_PROVIDER: groq
-      HINDSIGHT_API_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
-      HINDSIGHT_API_LLM_MODEL: openai/gpt-oss-20b
-      HINDSIGHT_API_URL: http://localhost:8888
-      GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-      UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu
-
-    steps:
-    - uses: actions/checkout@v4
-
-    - name: Download CLI artifact
-      uses: actions/download-artifact@v4
-      with:
-        name: hindsight-cli
-        path: /tmp/cli
-
-    - name: Make CLI executable
-      run: chmod +x /tmp/cli/hindsight
-
     - name: Install uv
       uses: astral-sh/setup-uv@v5
       with:
@@ -222,7 +225,7 @@ jobs:
 
     - name: Install API dependencies
       working-directory: ./hindsight-api
-      run: uv sync --no-install-project --index-strategy unsafe-best-match
+      run: uv sync --frozen --no-install-project --index-strategy unsafe-best-match
 
     - name: Create .env file
       run: |
@@ -251,7 +254,7 @@ jobs:
 
     - name: Run CLI smoke test
       run: |
-        HINDSIGHT_CLI=/tmp/cli/hindsight ./hindsight-cli/smoke-test.sh
+        HINDSIGHT_CLI=hindsight-cli/target/release/hindsight ./hindsight-cli/smoke-test.sh
 
     - name: Show API server logs
       if: always()
@@ -274,16 +277,35 @@ jobs:
       run: helm lint helm/hindsight
 
   build-docker-images:
+    name: Build Docker (${{ matrix.name }})
     runs-on: ubuntu-latest
     strategy:
       matrix:
         include:
           - target: api-only
             name: api
+            variant: full
+            build_args: ""
+          - target: api-only
+            name: api-slim
+            variant: slim
+            build_args: |
+              INCLUDE_LOCAL_MODELS=false
+              PRELOAD_ML_MODELS=false
           - target: cp-only
             name: control-plane
+            variant: full
+            build_args: ""
           - target: standalone
             name: standalone
+            variant: full
+            build_args: ""
+          - target: standalone
+            name: standalone-slim
+            variant: slim
+            build_args: |
+              INCLUDE_LOCAL_MODELS=false
+              PRELOAD_ML_MODELS=false
 
     steps:
     - uses: actions/checkout@v4
@@ -302,20 +324,31 @@ jobs:
     - name: Set up Docker Buildx
       uses: docker/setup-buildx-action@v3
 
-    - name: Build ${{ matrix.name }} image
+    - name: Build ${{ matrix.name }} image (${{ matrix.variant }})
       uses: docker/build-push-action@v6
       with:
         context: .
         file: docker/standalone/Dockerfile
         target: ${{ matrix.target }}
+        build-args: ${{ matrix.build_args }}
         push: false
-        load: false
+        load: ${{ matrix.variant == 'slim' }}
+        tags: hindsight-${{ matrix.name }}:test
+        # Removed GitHub Actions cache (type=gha) - it frequently returns 502 errors
+        # causing buildx to fail with "failed to parse error response 502"
+        # Build will be slower but more reliable
 
-    # TODO: Re-enable smoke test when disk space issue is resolved
-    # - name: Smoke test - verify container starts
-    #   env:
-    #     GROQ_API_KEY: ${{ secrets.GROQ_API_KEY }}
-    #   run: ./scripts/docker-smoke-test.sh "hindsight-${{ matrix.name }}:test" "${{ matrix.target }}"
+    # Only test slim variants to save disk space (they're much smaller)
+    # Slim variants require external embedding providers
+    - name: Smoke test - verify container starts
+      if: matrix.variant == 'slim'
+      env:
+        GROQ_API_KEY: ${{ secrets.GROQ_API_KEY }}
+        HINDSIGHT_API_EMBEDDINGS_PROVIDER: openai
+        HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
+        HINDSIGHT_API_RERANKER_PROVIDER: cohere
+        HINDSIGHT_API_COHERE_API_KEY: ${{ secrets.COHERE_API_KEY }}
+      run: ./docker/test-image.sh "hindsight-${{ matrix.name }}:test" "${{ matrix.target }}"
 
   test-api:
     runs-on: ubuntu-latest
@@ -352,7 +385,7 @@ jobs:
 
     - name: Install dependencies
       working-directory: ./hindsight-api
-      run: uv sync --extra test --no-install-project --index-strategy unsafe-best-match
+      run: uv sync --frozen --extra test --no-install-project --index-strategy unsafe-best-match
 
     - name: Cache HuggingFace models
       uses: actions/cache@v4
@@ -413,11 +446,11 @@ jobs:
 
     - name: Install client test dependencies
       working-directory: ./hindsight-clients/python
-      run: uv sync --extra test --index-strategy unsafe-best-match
+      run: uv sync --frozen --extra test --index-strategy unsafe-best-match
 
     - name: Install API dependencies
       working-directory: ./hindsight-api
-      run: uv sync --no-install-project --index-strategy unsafe-best-match
+      run: uv sync --frozen --no-install-project --index-strategy unsafe-best-match
 
     - name: Create .env file
       run: |
@@ -490,7 +523,7 @@ jobs:
 
     - name: Install API dependencies
       working-directory: ./hindsight-api
-      run: uv sync --no-install-project --index-strategy unsafe-best-match
+      run: uv sync --frozen --no-install-project --index-strategy unsafe-best-match
 
     - name: Install TypeScript client dependencies
       working-directory: ./hindsight-clients/typescript
@@ -578,7 +611,7 @@ jobs:
 
     - name: Install API dependencies
       working-directory: ./hindsight-api
-      run: uv sync --no-install-project --index-strategy unsafe-best-match
+      run: uv sync --frozen --no-install-project --index-strategy unsafe-best-match
 
     - name: Create .env file
       run: |
@@ -615,6 +648,80 @@ jobs:
         echo "=== API Server Logs ==="
         cat /tmp/api-server.log || echo "No API server log found"
 
+  test-go-client:
+    runs-on: ubuntu-latest
+    env:
+      HINDSIGHT_API_LLM_PROVIDER: groq
+      HINDSIGHT_API_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
+      HINDSIGHT_API_LLM_MODEL: openai/gpt-oss-20b
+      HINDSIGHT_API_URL: http://localhost:8888
+      GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+      # Prefer CPU-only PyTorch in CI (but keep PyPI for everything else)
+      UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu
+
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Install uv
+      uses: astral-sh/setup-uv@v5
+      with:
+        enable-cache: true
+        prune-cache: false
+
+    - name: Set up Python
+      uses: actions/setup-python@v5
+      with:
+        python-version-file: ".python-version"
+
+    - name: Set up Go
+      uses: actions/setup-go@v5
+      with:
+        go-version: '1.23'
+        cache-dependency-path: hindsight-clients/go/go.sum
+
+    - name: Build API
+      working-directory: ./hindsight-api
+      run: uv build
+
+    - name: Install API dependencies
+      working-directory: ./hindsight-api
+      run: uv sync --frozen --no-install-project --index-strategy unsafe-best-match
+
+    - name: Create .env file
+      run: |
+        cat > .env << EOF
+        HINDSIGHT_API_LLM_PROVIDER=${{ env.HINDSIGHT_API_LLM_PROVIDER }}
+        HINDSIGHT_API_LLM_API_KEY=${{ env.HINDSIGHT_API_LLM_API_KEY }}
+        HINDSIGHT_API_LLM_MODEL=${{ env.HINDSIGHT_API_LLM_MODEL }}
+        EOF
+
+    - name: Start API server
+      run: |
+        ./scripts/dev/start-api.sh > /tmp/api-server.log 2>&1 &
+        echo "Waiting for API server to be ready..."
+        for i in {1..60}; do
+          if curl -sf http://localhost:8888/health > /dev/null 2>&1; then
+            echo "API server is ready after ${i}s"
+            break
+          fi
+          if [ $i -eq 60 ]; then
+            echo "API server failed to start after 60s"
+            cat /tmp/api-server.log
+            exit 1
+          fi
+          sleep 1
+        done
+
+    - name: Run Go client tests
+      working-directory: ./hindsight-clients/go
+      run: go test -v -tags=integration
+
+    - name: Show API server logs
+      if: always()
+      run: |
+        echo "=== API Server Logs ==="
+        cat /tmp/api-server.log || echo "No API server log found"
+
   test-integration:
     runs-on: ubuntu-latest
     env:
@@ -645,11 +752,11 @@ jobs:
 
     - name: Install API dependencies
       working-directory: ./hindsight-api
-      run: uv sync --no-install-project --index-strategy unsafe-best-match
+      run: uv sync --frozen --no-install-project --index-strategy unsafe-best-match
 
     - name: Install integration test dependencies
       working-directory: ./hindsight-integration-tests
-      run: uv sync
+      run: uv sync --frozen
 
     - name: Cache HuggingFace models
       uses: actions/cache@v4
@@ -729,7 +836,7 @@ jobs:
 
     - name: Install dependencies
       working-directory: ./hindsight-integrations/litellm
-      run: uv sync --extra dev
+      run: uv sync --frozen --extra dev
 
     - name: Run tests
       working-directory: ./hindsight-integrations/litellm
@@ -738,9 +845,9 @@ jobs:
   test-embed:
     runs-on: ubuntu-latest
     env:
-      HINDSIGHT_EMBED_LLM_PROVIDER: groq
-      HINDSIGHT_EMBED_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
-      HINDSIGHT_EMBED_LLM_MODEL: openai/gpt-oss-20b
+      HINDSIGHT_API_LLM_PROVIDER: groq
+      HINDSIGHT_API_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
+      HINDSIGHT_API_LLM_MODEL: openai/gpt-oss-20b
       # Prefer CPU-only PyTorch in CI
       UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu
 
@@ -760,7 +867,7 @@ jobs:
 
     - name: Install dependencies
       working-directory: ./hindsight-embed
-      run: uv sync --index-strategy unsafe-best-match
+      run: uv sync --frozen --index-strategy unsafe-best-match
 
     - name: Cache HuggingFace models
       uses: actions/cache@v4
@@ -771,13 +878,65 @@ jobs:
           ${{ runner.os }}-huggingface-embed-
           ${{ runner.os }}-huggingface-
 
+    - name: Run unit and integration tests
+      working-directory: ./hindsight-embed
+      run: uv run pytest tests/ -v
+
     - name: Run smoke test
       working-directory: ./hindsight-embed
       run: ./test.sh
 
+  test-hindsight-all:
+    runs-on: ubuntu-latest
+    env:
+      HINDSIGHT_API_LLM_PROVIDER: groq
+      HINDSIGHT_API_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
+      HINDSIGHT_API_LLM_MODEL: openai/gpt-oss-20b
+      # For test_server_integration.py compatibility
+      HINDSIGHT_LLM_PROVIDER: groq
+      HINDSIGHT_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
+      HINDSIGHT_LLM_MODEL: openai/gpt-oss-20b
+      # Prefer CPU-only PyTorch in CI
+      UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu
+
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Install uv
+      uses: astral-sh/setup-uv@v5
+      with:
+        enable-cache: true
+        prune-cache: false
+
+    - name: Set up Python
+      uses: actions/setup-python@v5
+      with:
+        python-version-file: ".python-version"
+
+    - name: Build hindsight-all
+      working-directory: ./hindsight
+      run: uv build
+
+    - name: Install dependencies
+      working-directory: ./hindsight
+      run: uv sync --frozen --extra test --index-strategy unsafe-best-match
+
+    - name: Cache HuggingFace models
+      uses: actions/cache@v4
+      with:
+        path: ~/.cache/huggingface
+        key: ${{ runner.os }}-huggingface-all-${{ hashFiles('hindsight/pyproject.toml') }}
+        restore-keys: |
+          ${{ runner.os }}-huggingface-all-
+          ${{ runner.os }}-huggingface-
+
+    - name: Run unit tests
+      working-directory: ./hindsight
+      run: uv run pytest tests/ -v
+
   test-doc-examples:
     runs-on: ubuntu-latest
-    needs: build-rust-cli
+    needs: test-rust-cli
     env:
       HINDSIGHT_API_LLM_PROVIDER: groq
       HINDSIGHT_API_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
@@ -820,11 +979,11 @@ jobs:
       working-directory: ./hindsight-api
       run: |
         uv build
-        uv sync --no-install-project --index-strategy unsafe-best-match
+        uv sync --frozen --no-install-project --index-strategy unsafe-best-match
 
     - name: Install Python client dependencies
       working-directory: ./hindsight-clients/python
-      run: uv sync --extra test --index-strategy unsafe-best-match
+      run: uv sync --frozen --extra test --index-strategy unsafe-best-match
 
     - name: Install TypeScript client
       run: |
@@ -856,30 +1015,11 @@ jobs:
           sleep 1
         done
 
-    - name: Run Python doc examples
-      working-directory: ./hindsight-clients/python
-      run: |
-        for f in ../../hindsight-docs/examples/api/*.py; do
-          echo "Running $f..."
-          uv run python "$f"
-        done
-
-    - name: Run Node.js doc examples
-      run: |
-        for f in hindsight-docs/examples/api/*.mjs; do
-          echo "Running $f..."
-          node "$f"
-        done
-
     - name: Configure CLI
       run: hindsight configure --api-url http://localhost:8888
 
-    - name: Run CLI doc examples
-      run: |
-        for f in hindsight-docs/examples/api/*.sh; do
-          echo "Running $f..."
-          bash "$f"
-        done
+    - name: Run all doc examples
+      run: ./scripts/test-doc-examples.sh
 
     - name: Show API server logs
       if: always()
@@ -887,6 +1027,78 @@ jobs:
         echo "=== API Server Logs ==="
         cat /tmp/api-server.log || echo "No API server log found"
 
+  test-upgrade:
+    runs-on: ubuntu-latest
+    env:
+      HINDSIGHT_API_LLM_PROVIDER: groq
+      HINDSIGHT_API_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
+      HINDSIGHT_API_LLM_MODEL: openai/gpt-oss-20b
+      GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+      UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu
+
+    steps:
+    - uses: actions/checkout@v4
+      with:
+        fetch-depth: 0  # Full history needed for git clone of tags
+
+    - name: Fetch tags
+      run: git fetch --tags
+
+    - name: Install uv
+      uses: astral-sh/setup-uv@v5
+      with:
+        enable-cache: true
+        prune-cache: false
+
+    - name: Set up Python
+      uses: actions/setup-python@v5
+      with:
+        python-version-file: ".python-version"
+
+    - name: Cache HuggingFace models
+      uses: actions/cache@v4
+      with:
+        path: ~/.cache/huggingface
+        key: ${{ runner.os }}-huggingface-${{ hashFiles('hindsight-api/pyproject.toml') }}
+        restore-keys: |
+          ${{ runner.os }}-huggingface-
+
+    - name: Install hindsight-dev dependencies
+      working-directory: ./hindsight-dev
+      run: uv sync --frozen --extra test --index-strategy unsafe-best-match
+
+    - name: Install current hindsight-api
+      working-directory: ./hindsight-api
+      run: uv sync --frozen --index-strategy unsafe-best-match
+
+    - name: Pre-download models
+      working-directory: ./hindsight-api
+      run: |
+        uv run python -c "
+        from sentence_transformers import SentenceTransformer, CrossEncoder
+        print('Downloading embedding model...')
+        SentenceTransformer('BAAI/bge-small-en-v1.5')
+        print('Downloading cross-encoder model...')
+        CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')
+        print('Models downloaded successfully')
+        "
+
+    - name: Run upgrade tests
+      working-directory: ./hindsight-dev
+      run: uv run pytest upgrade_tests/ -v --tb=short
+
+    - name: Show upgrade test logs
+      if: always()
+      run: |
+        echo "=== Upgrade Test Server Logs ==="
+        for log in /tmp/upgrade-test-*.log; do
+          if [ -f "$log" ]; then
+            echo ""
+            echo "--- $log ---"
+            tail -500 "$log"
+          fi
+        done
+
   verify-generated-files:
     runs-on: ubuntu-latest
     env:
@@ -928,9 +1140,9 @@ jobs:
 
     - name: Install Python dependencies
       run: |
-        cd hindsight-dev && uv sync --index-strategy unsafe-best-match
-        cd ../hindsight-api && uv sync --index-strategy unsafe-best-match
-        cd ../hindsight-embed && uv sync --index-strategy unsafe-best-match
+        cd hindsight-dev && uv sync --frozen --index-strategy unsafe-best-match
+        cd ../hindsight-api && uv sync --frozen --index-strategy unsafe-best-match
+        cd ../hindsight-embed && uv sync --frozen --index-strategy unsafe-best-match
 
     - name: Run generate-openapi
       run: ./scripts/generate-openapi.sh
@@ -957,4 +1169,52 @@ jobs:
           git diff --stat
           exit 1
         fi
-        echo "✓ All generated files are up to date"
+        echo "✓ All generated files are up to date"
+
+  check-openapi-compatibility:
+    runs-on: ubuntu-latest
+    env:
+      UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu
+
+    steps:
+    - uses: actions/checkout@v4
+      with:
+        fetch-depth: 0  # Fetch full git history to access base branch
+
+    - name: Install uv
+      uses: astral-sh/setup-uv@v5
+      with:
+        enable-cache: true
+
+    - name: Set up Python
+      uses: actions/setup-python@v5
+      with:
+        python-version-file: ".python-version"
+
+    - name: Install hindsight-dev dependencies
+      run: |
+        cd hindsight-dev && uv sync --frozen --index-strategy unsafe-best-match
+
+    - name: Check OpenAPI compatibility with base branch
+      run: |
+        # Get the base branch (usually main)
+        BASE_BRANCH="${{ github.base_ref }}"
+
+        if [ -z "$BASE_BRANCH" ]; then
+          echo "⚠️  Warning: No base branch found (not a PR?). Skipping compatibility check."
+          exit 0
+        fi
+
+        echo "Checking OpenAPI compatibility against base branch: $BASE_BRANCH"
+
+        # Extract the old OpenAPI spec from base branch
+        git show "origin/$BASE_BRANCH:hindsight-docs/static/openapi.json" > /tmp/old-openapi.json
+
+        if [ ! -s /tmp/old-openapi.json ]; then
+          echo "⚠️  Warning: Could not find OpenAPI spec in base branch. Skipping compatibility check."
+          exit 0
+        fi
+
+        # Check compatibility using our tool
+        cd hindsight-dev
+        uv run check-openapi-compatibility /tmp/old-openapi.json ../hindsight-docs/static/openapi.json

.gitignore

@@ -29,6 +29,7 @@ nltk_data/
 
 # Monitoring stack (Prometheus/Grafana binaries and data)
 .monitoring/
+.pgbouncer/
 
 # Large benchmark datasets (will be downloaded automatically)
 **/longmemeval_s_cleaned.json
@@ -44,9 +45,15 @@ hindsight-docs/static/llms-full.txt
 
 hindsight-dev/benchmarks/locomo/results/
 hindsight-dev/benchmarks/longmemeval/results/
+hindsight-dev/benchmarks/consolidation/results/
+hindsight-dev/benchmarks/perf/results/
+benchmarks/results/
 hindsight-cli/target
 hindsight-clients/rust/target
 .claude
 whats-next.md
 TASK.md
-CHANGELOG.md
+# Changelog is now tracked in hindsight-docs/src/pages/changelog.md
+# CHANGELOG.md
+
+blog-post*

AGENTS.md

@@ -1,153 +1,3 @@
 # AGENTS.md
 
-This document captures architectural decisions and coding conventions for the Hindsight project.
-
-## Documentation
-
-- **Main documentation**: [hindsight-docs/docs/developer/](./hindsight-docs/docs/developer/)
-- **Use case patterns**: [hindsight-docs/docs/cookbook/](./hindsight-docs/docs/cookbook/)
-- **API reference**: Auto-generated from OpenAPI spec
-
-## Project Structure
-
-```
-hindsight/              # Python package for embedded usage
-hindsight-api/          # FastAPI server (core memory engine)
-hindsight-cli/          # Rust CLI client
-hindsight-embed/        # Embedded CLI (no server needed)
-hindsight-control-plane/ # Next.js admin UI
-hindsight-docs/         # Docusaurus documentation site
-hindsight-dev/          # Development tools and benchmarks
-hindsight-integrations/ # Framework integrations (LangChain, etc.)
-hindsight-clients/      # Generated API clients (Python, TypeScript, Rust)
-```
-
-## Core Concepts
-
-### Memory Banks
-- Each bank is an isolated memory store (like a "brain" for one user/agent)
-- Banks contain: memory units (facts), entities, documents, entity links
-- Banks have a **disposition** (personality traits) and **background** (context)
-- Bank isolation is strict - no cross-bank data leakage
-
-### Memory Types
-- **World facts**: General knowledge ("The sky is blue")
-- **Experience facts**: Personal experiences ("I visited Paris in 2023")
-- **Opinion facts**: Beliefs with confidence scores ("Paris is beautiful" - 0.9 confidence)
-
-### Operations
-- **Retain**: Store new memories (extracts facts, entities, relationships)
-- **Recall**: Retrieve memories (semantic, BM25, graph, temporal search)
-- **Reflect**: Deep analysis to form new insights/opinions
-
-## API Design Decisions
-
-### Single Bank Per Request
-- All API endpoints (`recall`, `reflect`, `retain`) operate on a single bank
-- Multi-bank queries are the **client/agent's responsibility** to orchestrate
-- This keeps the API simple and the isolation model clear
-
-### Disposition Traits (3-trait system)
-- **Skepticism** (1-5): How skeptical vs trusting when forming opinions
-- **Literalism** (1-5): How literally to interpret information
-- **Empathy** (1-5): How much to consider emotional context
-- These influence the `reflect` operation, not `recall`
-- Background info also only affects `reflect` (opinion formation)
-
-## Multi-Bank Architecture Patterns
-
-See [hindsight-docs/docs/cookbook/](./hindsight-docs/docs/cookbook/) for detailed guides:
-
-- **Per-User Memory**: One bank per user, simplest pattern
-- **Support Agent + Shared Knowledge**: User bank + shared docs bank, client orchestrates
-
-## Developer Guide
-
-### Running the API Server
-
-```bash
-# From project root
-./scripts/dev/start-api.sh
-
-# With options
-./scripts/dev/start-api.sh --reload --port 8888 --log-level debug
-```
-
-### Running Tests
-
-```bash
-# API tests
-cd hindsight-api
-uv run pytest tests/
-
-# Specific test
-uv run pytest tests/test_http_api_integration.py -v
-```
-
-### Generating OpenAPI Spec
-
-After changing API endpoints, regenerate the OpenAPI spec and docs:
-
-```bash
-./scripts/generate-openapi.sh
-```
-
-This will:
-1. Generate `openapi.json` at project root
-2. Copy to `hindsight-docs/openapi.json`
-3. Regenerate API reference documentation
-
-### Generating API Clients
-
-After updating the OpenAPI spec, regenerate all clients:
-
-```bash
-./scripts/generate-clients.sh
-```
-
-This generates:
-- **Rust client**: `hindsight-clients/rust/` (via progenitor in build.rs)
-- **Python client**: `hindsight-clients/python/` (via openapi-generator Docker)
-- **TypeScript client**: `hindsight-clients/typescript/` (via @hey-api/openapi-ts)
-
-Note: The maintained wrapper `hindsight_client.py` and `README.md` are preserved during regeneration.
-
-### Running the Documentation Site
-
-```bash
-./scripts/dev/start-docs.sh
-```
-
-### Running the Control Plane
-
-```bash
-./scripts/dev/start-control-plane.sh
-```
-
-## Code Style
-
-### Python (hindsight-api)
-- Use `uv` for package management
-- Async throughout (asyncpg, async FastAPI endpoints)
-- Pydantic models for request/response validation
-- No py files at project root - maintain clean directory structure
-
-### TypeScript (control-plane, clients)
-- Next.js with App Router for control plane
-- Tailwind CSS with shadcn/ui components
-
-### Rust (CLI)
-- Async with tokio
-- reqwest for HTTP client
-- progenitor for API client generation
-
-## Database
-
-- PostgreSQL with pgvector extension
-- Schema managed via Alembic migrations in `hindsight-api/alembic/`, db migrations happen during api startup, no manual commands
-- Key tables: `banks`, `memory_units`, `documents`, `entities`, `entity_links`
-
-# Branding
-## Colors
-- Primary: gradient from #0074d9 to #009296
-
+See [CLAUDE.md](./CLAUDE.md) for project documentation and coding conventions.

CLAUDE.md

@@ -7,8 +7,7 @@ This file provides guidance to Claude Code (claude.ai/code) when working with co
 Hindsight is an agent memory system that provides long-term memory for AI agents using biomimetic data structures. Memories are organized as:
 - **World facts**: General knowledge ("The sky is blue")
 - **Experience facts**: Personal experiences ("I visited Paris in 2023")
-- **Opinion facts**: Beliefs with confidence scores ("Paris is beautiful" - 0.9 confidence)
-- **Observations**: Complex mental models derived from reflection
+- **Mental models**: Consolidated knowledge synthesized from facts ("User prefers functional programming patterns")
 
 ## Development Commands
 
@@ -46,6 +45,7 @@ cd hindsight-control-plane && npm run dev
 ./scripts/dev/start-docs.sh
 ```
 
+
 ### Generating Clients/OpenAPI
 ```bash
 # Regenerate OpenAPI spec after API changes (REQUIRED after changing endpoints)
@@ -57,8 +57,15 @@ cd hindsight-control-plane && npm run dev
 
 ### Benchmarks
 ```bash
+# Accuracy benchmarks
 ./scripts/benchmarks/run-longmemeval.sh
 ./scripts/benchmarks/run-locomo.sh
+
+# Performance benchmarks
+./scripts/benchmarks/run-consolidation.sh
+./scripts/benchmarks/run-retain-perf.sh --document <path>  # Requires API server running
+
+# Results viewer
 ./scripts/benchmarks/start-visualizer.sh  # View results at localhost:8001
 ```
 
@@ -101,13 +108,59 @@ cd hindsight-control-plane && npm run dev
 Main operations:
 - **Retain**: Store memories, extracts facts/entities/relationships
 - **Recall**: Retrieve memories via 4 parallel strategies (semantic, BM25, graph, temporal) + reranking
-- **Reflect**: Deep analysis forming new opinions/observations (disposition-aware)
+- **Reflect**: Disposition-aware reasoning using memories and mental models.
 
 ### Database
 PostgreSQL with pgvector. Schema managed via Alembic migrations in `hindsight-api/hindsight_api/alembic/`. Migrations run automatically on API startup.
 
 Key tables: `banks`, `memory_units`, `documents`, `entities`, `entity_links`
 
+### Adding Database Migrations
+
+1. **Create a new migration file** in `hindsight-api/hindsight_api/alembic/versions/`:
+   - File name format: `<revision_id>_<description>.py` (e.g., `f1a2b3c4d5e6_add_new_index.py`)
+   - Use a unique hex revision ID (12 chars)
+   - Set `down_revision` to the previous migration's revision ID
+
+2. **Migration template**:
+   ```python
+   """Description of the migration
+
+   Revision ID: f1a2b3c4d5e6
+   Revises: <previous_revision_id>
+   Create Date: YYYY-MM-DD
+   """
+   from collections.abc import Sequence
+   from alembic import context, op
+
+   revision: str = "f1a2b3c4d5e6"
+   down_revision: str | Sequence[str] | None = "<previous_revision_id>"
+   branch_labels: str | Sequence[str] | None = None
+   depends_on: str | Sequence[str] | None = None
+
+   def _get_schema_prefix() -> str:
+       """Get schema prefix for table names (required for multi-tenant support)."""
+       schema = context.config.get_main_option("target_schema")
+       return f'"{schema}".' if schema else ""
+
+   def upgrade() -> None:
+       schema = _get_schema_prefix()
+       op.execute(f"CREATE INDEX ... ON {schema}table_name(...)")
+
+   def downgrade() -> None:
+       schema = _get_schema_prefix()
+       op.execute(f"DROP INDEX IF EXISTS {schema}index_name")
+   ```
+
+3. **Run migrations locally**:
+   ```bash
+   # Set database URL and run migrations
+   uv run hindsight-admin run-db-migration
+
+   # Run on a specific tenant schema
+   uv run hindsight-admin run-db-migration --schema tenant_xyz
+   ```
+
 ## Key Conventions
 
 ### Code Quality
@@ -128,12 +181,63 @@ This runs the same checks as the pre-commit hook (Ruff for Python, ESLint/Pretti
 - Multi-bank queries are client responsibility to orchestrate
 - Disposition traits only affect reflect, not recall
 
+### Control Plane API Routes
+
+When adding or modifying parameters in the dataplane API (hindsight-api), you must also update the control plane routes that proxy to it:
+
+1. **API Routes** (`hindsight-control-plane/src/app/api/`):
+   - `recall/route.ts` - proxies to `/v1/default/banks/{bank_id}/memories/recall`
+   - `reflect/route.ts` - proxies to `/v1/default/banks/{bank_id}/reflect`
+   - `memories/retain/route.ts` - proxies to `/v1/default/banks/{bank_id}/memories/retain`
+   - Other routes follow the same pattern
+
+2. **Client types** (`hindsight-control-plane/src/lib/api.ts`):
+   - Update the TypeScript type definitions for `recall()`, `reflect()`, `retain()` etc.
+
+3. **Checklist when adding new API parameters**:
+   - Add parameter extraction in the route handler (destructure from `body`)
+   - Pass the parameter to the SDK call
+   - Update the client type definition in `lib/api.ts`
+   - Update any UI components that need to use the new parameter
+
 ### Python Style
 - Python 3.11+, type hints required
 - Async throughout (asyncpg, async FastAPI)
 - Pydantic models for request/response
 - Ruff for linting (line-length 120)
 - No Python files at project root - maintain clean directory structure
+- **Never use multi-item tuple return values** - prefer dataclass or Pydantic model for structured returns
+
+### Type Safety with Pydantic Models
+**NEVER use raw `dict` types for structured data.** Always use Pydantic models:
+- Use Pydantic `BaseModel` for all data structures passed between functions
+- Add `@field_validator` for type coercion (e.g., ensuring datetimes are timezone-aware)
+- Avoid `dict.get()` patterns - use typed model attributes instead
+- Parse external data (JSON, API responses) into Pydantic models at the boundary
+- This catches type errors at parse time, not deep in business logic
+
+```python
+# BAD - error-prone dict access
+def process(data: dict) -> str:
+    return data.get("name", "")  # No validation, silent failures
+
+# GOOD - typed and validated
+class UserData(BaseModel):
+    name: str
+    created_at: datetime
+
+    @field_validator("created_at", mode="before")
+    @classmethod
+    def ensure_tz_aware(cls, v):
+        if isinstance(v, str):
+            v = datetime.fromisoformat(v.replace("Z", "+00:00"))
+        if v.tzinfo is None:
+            return v.replace(tzinfo=timezone.utc)
+        return v
+
+def process(data: UserData) -> str:
+    return data.name  # Type-safe, validated at construction
+```
 
 ### TypeScript Style
 - Next.js App Router for control plane
@@ -141,26 +245,61 @@ This runs the same checks as the pre-commit hook (Ruff for Python, ESLint/Pretti
 
 ### Adding New API Configuration Flags
 
-When adding a new environment variable configuration:
+Configuration follows a hierarchical system: **Global (env vars) → Tenant (via extension) → Bank (database)**.
+
+Fields must be categorized as either **hierarchical** (can be overridden per-tenant/bank) or **static** (server-level only).
+
+#### Adding a New Configuration Field
 
 1. **config.py** (`hindsight-api/hindsight_api/config.py`):
-   - Add `ENV_*` constant for the environment variable name
+   - Add `ENV_*` constant for the environment variable name (e.g., `ENV_MY_SETTING = "HINDSIGHT_API_MY_SETTING"`)
    - Add `DEFAULT_*` constant for the default value
-   - Add field to `HindsightConfig` dataclass
+   - Add field to `HindsightConfig` dataclass with type annotation
+   - **Mark as hierarchical or static** by adding to `_HIERARCHICAL_FIELDS` set (hierarchical) or leaving it out (static)
    - Add initialization in `from_env()` method
 
+   ```python
+   # Hierarchical field (can be overridden per-bank)
+   _HIERARCHICAL_FIELDS = {
+       ...,
+       "my_setting",  # Add here for hierarchical
+   }
+
+   # Static field - just don't add to _HIERARCHICAL_FIELDS
+   ```
+
 2. **main.py** (`hindsight-api/hindsight_api/main.py`):
    - Add field to the manual `HindsightConfig()` constructor call (search for "CLI override")
 
-3. **Use the config** in code:
+3. **Use hierarchical config in MemoryEngine**:
+   ```python
+   # Config is resolved automatically per bank via ConfigResolver
+   config_dict = await self._config_resolver.get_bank_config(bank_id, context)
+   value = config_dict["my_setting"]
+   ```
+
+4. **Use static config** (non-hierarchical):
    ```python
    from ...config import get_config
    config = get_config()
-   value = config.your_new_field
+   value = config.my_static_field
    ```
 
-4. **Documentation** (`hindsight-docs/docs/developer/configuration.md`):
+5. **Documentation** (`hindsight-docs/docs/developer/configuration.md`):
    - Add to appropriate section table with Variable, Description, Default
+   - Mark if it's hierarchical (can be overridden per-bank)
+
+#### Hierarchical vs Static Guidelines
+
+**Hierarchical** (per-bank overridable):
+- LLM settings (provider, model, API key, base URL)
+- Operation-specific settings (retain mode, chunk size, etc.)
+- Feature flags that vary by customer/bank
+
+**Static** (server-level only):
+- Infrastructure settings (database URL, port, host)
+- Global limits (max concurrent operations)
+- System-wide feature flags
 
 ## Environment Setup
 
@@ -184,3 +323,4 @@ Optional (uses local models by default):
 - `HINDSIGHT_API_EMBEDDINGS_PROVIDER`: local (default) or tei
 - `HINDSIGHT_API_RERANKER_PROVIDER`: local (default) or tei
 - `HINDSIGHT_API_DATABASE_URL`: External PostgreSQL (uses embedded pg0 by default)
+- `HINDSIGHT_API_ENABLE_BANK_CONFIG_API`: Enable per-bank config API (default: false, disabled for security)

CONTRIBUTING.md

@@ -93,6 +93,34 @@ uv run ty check hindsight_api  # Type check
 3. Run tests to ensure nothing breaks
 4. Submit a PR with a clear description of changes
 
+## Release Process
+
+The project uses `scripts/release.sh` for creating releases. This script automates the entire release workflow:
+
+1. Bumps version in all components (API, clients, CLI, control plane, Helm)
+2. **Regenerates OpenAPI spec and client SDKs** (Python, TypeScript, Rust)
+3. Updates documentation versioning
+4. Creates a commit and git tag
+5. Pushes to GitHub (triggers CI/CD to publish packages)
+
+### Usage
+
+```bash
+./scripts/release.sh <version>
+```
+
+**Example:**
+```bash
+./scripts/release.sh 0.5.0
+```
+
+### Important for Developers
+
+- During development, version bumps in `__init__.py` do NOT require client regeneration
+- Clients are only regenerated during releases
+- Do not manually run `./scripts/generate-clients.sh` unless testing generation changes
+- Client version comments will reflect the API version from the latest release
+
 ## Reporting Issues
 
 Open an issue on GitHub with:

README.md

@@ -1,11 +1,11 @@
 <div align="center">
 
-![Hindsight Banner](./hindsight-docs/static/img/banner.svg)
+![Hindsight Banner](./hindsight-docs/static/img/hindsight-github-banner.png)
 
-[Documentation](https://hindsight.vectorize.io) • [Paper](https://arxiv.org/abs/2512.12818) • [Cookbook](https://hindsight.vectorize.io/cookbook) • [Hindsight Cloud](https://vectorize.io/hindsight/cloud)
+[Documentation](https://hindsight.vectorize.io) • [Paper](https://arxiv.org/abs/2512.12818) • [Cookbook](https://hindsight.vectorize.io/cookbook) • [Hindsight Cloud](https://ui.hindsight.vectorize.io/signup)
 
 [![CI](https://github.com/vectorize-io/hindsight/actions/workflows/release.yml/badge.svg)](https://github.com/vectorize-io/hindsight/actions/workflows/release.yml)
-[![Slack Community](https://img.shields.io/badge/Slack-Join%20Community-4A154B?logo=slack)](https://join.slack.com/t/hindsight-space/shared_invite/zt-3klo21kua-VUCC_zHP5rIcXFB1_5yw6A)
+[![Slack Community](https://img.shields.io/badge/Slack-Join%20Community-4A154B?logo=slack)](https://join.slack.com/t/hindsight-space/shared_invite/zt-3nhbm4w29-LeSJ5Ixi6j8PdiYOCPlOgg)
 [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
 ![PyPI - Downloads](https://img.shields.io/pypi/dm/hindsight-api?label=PyPI)
 ![NPM Downloads](https://img.shields.io/npm/dm/%40vectorize-io%2Fhindsight-client?logoColor=orange&label=NPM&color=blue&link=https%3A%2F%2Fwww.npmjs.com%2Fpackage%2F%40vectorize-io%2Fhindsight-client)
@@ -17,76 +17,76 @@
 
 ## What is Hindsight?
 
-Hindsight™ is an agent memory system built to create smarter agents that learn over time. It eliminates the shortcomings of alternative techniques such as RAG and knowledge graph and delivers state-of-the-art performance on long term memory tasks.
+Hindsight™ is an agent memory system built to create smarter agents that learn over time. Most agent memory systems focus on recalling conversation history. Hindsight is focused on making agents that learn, not just remember.
 
-Hindsight addresses common challenges that have frustrated AI engineers building agents to automate tasks and assist users with conversational interfaces. Many of these challenges stem directly from a lack of memory.
 
-- **Inconsistency:** Agents complete tasks successfully one time, then fail when asked to complete the same task again. Memory gives the agent a mechanism to remember what worked and what didn't and to use that information to reduce errors and improve consistency.
-- **Hallucinations:** Long term memory can be seeded with external knowledge to ground agent behavior in reliable sources to augment training data.
-- **Cognitive Overload:** As workflows get complex, retrievals, tool calls, user messages and agent responses can grow to fill the context window leading to context rot. Short term memory optimization allows agents to reduce tokens and focus context by removing irrelevant details.
+<video src="https://github.com/user-attachments/assets/923b798d-3581-4897-bb62-9cfa5a931682" controls></video>
 
-## How is Hindsight Different From Other Memory Systems?
-
-![Overview](./hindsight-docs/static/img/hindsight-overview.webp)
-
-Most agent memory implementation rely on basic vector search or sometimes use a knowledge graph. Hindsight uses biomimetic data structures to organize agent memories in a way that is more like how human memory works:
-
-- **World:** Facts about the world ("The stove gets hot")
-- **Experiences:** Agent's own experiences ("I touched the stove and it really hurt")
-- **Opinion:** Beliefs with confidence scores ("I shouldn't touch the stove again" - .99 confidence)
-- **Observation:** Complex mental models derived by reflecting on facts and experiences ("Curling irons, ovens, and fire are also hot. I shouldn't touch those either.")
-
-Memories in Hindsight are stored in banks (i.e. memory banks). When memories are added to Hindsight, they are pushed into either the world facts or experiences memory pathway. They are then represented as a combination of entities, relationships, and time series with sparse/dense vector representations to aid in later recall.
-
-Hindsight provides three simple methods to interact with the system:
-
-- **Retain:** Provide information to Hindsight that you want it to remember
-- **Recall:** Retrieve memories from Hindsight
-- **Reflect:** Reflect on memories and experiences to generate new observations and insights from existing memories.
-
-### Agent Memory That Learns
-
-A key goal of Hindsight is to build agent memory that enables agents to learn and improve over time. This is the role of the `reflect` operation which provides the agent to form broader opinions and observations over time.
-
-For example, imagine a product support agent that is helping a user troubleshoot a problem. It uses a `search-documentation` tool it found on an MCP server. Later in the conversation, the agent discovers that the documentation returned from the tool wasn't for the product the user was asking about. The agent now has an experience in its memory bank. And just like humans, we want that agent to learn from its experience.
-
-As the agent gains more experiences, `reflect` allows the agent to form observations about what worked, what didn't, and what to do differently the next time it encounters a similar task.
-
----
+It eliminates the shortcomings of alternative techniques such as RAG and knowledge graph and delivers state-of-the-art performance on long term memory tasks.
 
 ## Memory Performance & Accuracy
 
-Hindsight has achieved state-of-the-art performance on the LongMemEval benchmark, widely used to assess memory system performance across a variety of conversational
-AI scenarios. The current reported performance of Hindsight and other agent memory solutions as of December 2025 is shown here:
+Hindsight is the most accurate agent memory system ever tested according to benchmark performance. It has achieved state-of-the-art performance on the LongMemEval benchmark, widely used to assess memory system performance across a variety of conversational AI scenarios. The current reported performance of Hindsight and other agent memory solutions as of January 2026 is shown here:
 
 ![Overview](./hindsight-docs/static/img/hindsight-bench.jpg)
 
-The benchmark performance data for Hindsight and GPT-4o (full context) have been reproduced by research collaborators at the Virginia Tech [Sanghani Center for Artificial Intelligence and Data Analytics](https://sanghani.cs.vt.edu/) and The Washington Post. Other scores are self-reported by software vendors.
+The benchmark performance data for Hindsight has been independently reproduced by research collaborators at the Virginia Tech [Sanghani Center for Artificial Intelligence and Data Analytics](https://sanghani.cs.vt.edu/) and The Washington Post. Other scores are self-reported by software vendors.
 
-A thorough examination of the techniques implemented in Hindsight and detailed breakdowns of benchmark performance are [available on arXiv](https://arxiv.org/abs/2512.12818). This research is currently being prepared for conference submission and the wider peer review process.
+Hindsight is being used in production at Fortune 500 enterprises and by a growing number of AI startups. 
+
+## Adding Hindsight to Your AI Agents
+
+The easiest way use Hindsight with an existing agent is with the LLM Wrapper. You can add memory to your agent with 2 lines of code. That will swap your current LLM client out with the Hindsight wrapper. After that, memories will be stored and retrieved automatically as you make LLM calls.
+
+If you need more control over how and when your agent stores and recalls memories, there's also a simple API you can integrate with using the SDKs or directly via HTTP.
+
+![Hindsight Banner](./hindsight-docs/static/img/migration-code.png)
+
+---
+
+> 🤖 **Using a coding agent?** Install the Hindsight documentation skill for instant access to docs while you code:
+> ```bash
+> npx skills add https://github.com/vectorize-io/hindsight --skill hindsight-docs
+> ```
+> Works with Claude Code, Cursor, and other AI coding assistants.
+
+---
 
-The benchmark results from this research can be inspected in our [visual benchmark explorer](https://hindsight-benchmarks.vercel.app). As additional improvements are made to Hindsight, new benchmark data will be available for review using this same tool.
 
 ## Quick Start
 
 ### Docker (recommended)
 
 ```bash
-export OPENAI_API_KEY=your-key
+export OPENAI_API_KEY=sk-xxx
 
 docker run --rm -it --pull always -p 8888:8888 -p 9999:9999 \
   -e HINDSIGHT_API_LLM_API_KEY=$OPENAI_API_KEY \
-  -e HINDSIGHT_API_LLM_MODEL=o3-mini \
   -v $HOME/.hindsight-docker:/home/hindsight/.pg0 \
   ghcr.io/vectorize-io/hindsight:latest
 ```
 
+>API: http://localhost:8888
+>UI: http://localhost:9999
+
 You can modify the LLM provider by setting `HINDSIGHT_API_LLM_PROVIDER`. Valid options are `openai`, `anthropic`, `gemini`, `groq`, `ollama`, and `lmstudio`. The documentation provides more details on [supported models](https://hindsight.vectorize.io/developer/models).
 
-API: http://localhost:8888  
-UI: http://localhost:9999
 
-Install client:
+
+### Docker (external PostgreSQL)
+
+```bash
+export OPENAI_API_KEY=sk-xxx
+export HINDSIGHT_DB_PASSWORD=choose-a-password
+cd docker/docker-compose
+docker compose up 
+```
+
+
+>API: http://localhost:8888
+>UI: http://localhost:9999
+
+### Client
 
 ```bash
 pip install hindsight-client -U
@@ -94,7 +94,7 @@ pip install hindsight-client -U
 npm install @vectorize-io/hindsight-client
 ```
 
-Python example:
+#### Python
 
 ```python
 from hindsight_client import Hindsight
@@ -111,7 +111,29 @@ client.recall(bank_id="my-bank", query="What does Alice do?")
 client.reflect(bank_id="my-bank", query="Tell me about Alice")
 ```
 
-### Python (embedded, no Docker)
+#### Node.js / TypeScript
+
+```bash
+npm install @vectorize-io/hindsight-client
+```
+
+```javascript
+const { HindsightClient } = require('@vectorize-io/hindsight-client');
+
+const main = async () => {
+  const client = new HindsightClient({ baseUrl: 'http://localhost:8888' });
+
+  await client.retain('my-bank', 'Alice loves hiking in Yosemite');
+
+  const results = await client.recall('my-bank', 'What does Alice like?');
+  console.log(results);
+}
+
+main();
+```
+
+
+### Python Embedded (no server required)
 
 ```bash
 pip install hindsight-all -U
@@ -131,25 +153,48 @@ with HindsightServer(
     results = client.recall(bank_id="my-bank", query="Where does Alice work?")
 ```
 
-### Node.js / TypeScript
 
-```bash
-npm install @vectorize-io/hindsight-client
-```
+---
 
-```javascript
-const { HindsightClient } = require('@vectorize-io/hindsight-client');
+## Use Cases
 
-const client = new HindsightClient({ baseUrl: 'http://localhost:8888' });
 
-await client.retain('my-bank', 'Alice loves hiking in Yosemite');
-await client.recall('my-bank', 'What does Alice like?');
-```
+Hindsight is built to support conversational AI agents as well as agents that are intended to perform tasks autonomously. The ideal use case for Hindsight are agents that require a blend of these features such as AI employees that need to handle open-ended tasks, change behavior based on user feedback, and learn to perform complex tasks to automate work at a level that approximates a human work. Hindsight can be used with simple AI workflows like those built with n8n and other similar tools, but may be overkill for such applications.
+
+### Per-User Memories and Chat History
+
+One of the simpler use cases you can use Hindsight for is to personalize AI chatbots and other conversational agents by storing and recalling memories associated with individual users.
+
+The requirements for this use case usually look something like this:
+
+![Per-User Memories](./hindsight-docs/static/img/per-user-memory-requirements.png)
+
+<video src="https://github.com/user-attachments/assets/4805e8e1-e7d1-47c6-a4f8-2344a5ec8906" controls></video>
+
+Satisfying these requirements in Hindsight is straightforward. When new user inputs and tool calls are ingested into Hindsight using the retain operation, custom metadata can be used to enrich the new memories. Metadata provides a convenient way to isolate memories that need to be restricted to a given user. Once these are fed into the retain operation, any raw memories and mental models that get created can be filtered when retrieving relevant memories. 
+
+![Per-User Memories](./hindsight-docs/static/img/per-user-memory-howto.png)
 
 ---
 
 ## Architecture & Operations
 
+![Overview](./hindsight-docs/static/img/hindsight-overview.webp)
+
+Most agent memory implementation rely on basic vector search or sometimes use a knowledge graph. Hindsight uses biomimetic data structures to organize agent memories in a way that is more like how human memory works:
+
+- **World:** Facts about the world ("The stove gets hot")
+- **Experiences:** Agent's own experiences ("I touched the stove and it really hurt")
+- **Mental Models:** Learned understanding of the agent's world formed by reflecting on raw memories and experiences.
+
+Memories in Hindsight are stored in banks (i.e. memory banks). When memories are added to Hindsight, they are pushed into either the world facts or experiences memory pathway. They are then represented as a combination of entities, relationships, and time series with sparse/dense vector representations to aid in later recall.
+
+Hindsight provides three simple methods to interact with the system:
+
+- **Retain:** Provide information to Hindsight that you want it to remember
+- **Recall:** Retrieve memories from Hindsight
+- **Reflect:** Reflect on memories and experiences to generate new observations and insights from existing memories.
+
 ### Retain
 
 The `retain` operation is used to push new memories into Hindsight. It tells Hindsight to _retain_ the information you pass in as an input.
@@ -208,7 +253,7 @@ The final output is trimmed as needed to fit within the token limit.
 
 ### Reflect
 
-The reflect operation is used to perform a more thorough analysis of existing memories. This allows the agent to form new connections between memories which are then persisted as opinions and/or observations. When building agents, the reflect operation is a key capability to enable the agent to learn from its experiences. 
+The reflect operation is used to perform a more thorough analysis of existing memories. This allows the agent to form new connections between memories and build a more thorough understanding of its world.
 
 For example, the `reflect` operation can be used to support use cases such as:
 
@@ -242,7 +287,7 @@ client.reflect(bank_id="my-bank", query="What should I know about Alice?")
 - [CLI](https://hindsight.vectorize.io/sdks/cli)
 
 **Community:**
-- [Slack](https://join.slack.com/t/hindsight-space/shared_invite/zt-3klo21kua-VUCC_zHP5rIcXFB1_5yw6A)
+- [Slack](https://join.slack.com/t/hindsight-space/shared_invite/zt-3nhbm4w29-LeSJ5Ixi6j8PdiYOCPlOgg)
 - [GitHub Issues](https://github.com/vectorize-io/hindsight/issues)
 
 ---

docker/docker-compose/external-pg/docker-compose.yaml

@@ -0,0 +1,54 @@
+# Docker Compose file for Hindsight with PostgreSQL and pgvector
+#
+# Make sure to set the required environment variables before running:
+# - HINDSIGHT_DB_PASSWORD: Password for the PostgreSQL user
+# - Configure LLM provider variables as needed (see below in the hindsight service)
+#
+# Usage:
+# docker compose up -d
+#
+# Optional environment variables with defaults:
+# - HINDSIGHT_VERSION: Hindsight application version (default: latest)
+# - HINDSIGHT_DB_USER: PostgreSQL user (default: hindsight_user)
+# - HINDSIGHT_DB_NAME: PostgreSQL database name (default: hindsight_db)
+# - HINDSIGHT_DB_VERSION: PostgreSQL version (default: 18)
+
+services:
+  db:
+    # Use a PostgreSQL-Image with pgvector extension pre-installed
+    # see https://hub.docker.com/r/pgvector/pgvector 
+    image: pgvector/pgvector:pg${HINDSIGHT_DB_VERSION:-18}
+    container_name: hindsight-db
+    restart: always
+    # Expose PostgreSQL port
+    # ports:
+      # - "5432:5432"
+    environment:
+      POSTGRES_USER: ${HINDSIGHT_DB_USER:-hindsight_user}
+      POSTGRES_PASSWORD: ${HINDSIGHT_DB_PASSWORD:?Please set the HINDSIGHT_DB_PASSWORD env variable}
+      POSTGRES_DB: ${HINDSIGHT_DB_NAME:-hindsight_db}
+    volumes:
+      - pg_data:/var/lib/postgresql/${HINDSIGHT_DB_VERSION:-18}/docker
+    networks:
+      - hindsight-net
+
+  hindsight:
+    image: ghcr.io/vectorize-io/hindsight:${HINDSIGHT_VERSION:-latest}
+    container_name: hindsight-app
+    ports:
+      - "8888:8888"
+      - "9999:9999"
+    environment:
+      - HINDSIGHT_API_LLM_API_KEY=${OPENAI_API_KEY?Please set the OPENAI_API_KEY env variable}
+      - HINDSIGHT_API_DATABASE_URL=postgresql://${HINDSIGHT_DB_USER:-hindsight_user}:${HINDSIGHT_DB_PASSWORD:?Please set the HINDSIGHT_DB_PASSWORD env variable}@db:5432/${HINDSIGHT_DB_NAME:-hindsight_db}
+    depends_on:
+      - db
+    networks:
+      - hindsight-net
+
+networks:
+  hindsight-net:
+    driver: bridge
+
+volumes:
+  pg_data:

docker/docker-compose/nginx/README.md

@@ -0,0 +1,96 @@
+# Nginx Reverse Proxy with Custom Base Path
+
+Deploy Hindsight API under `/hindsight` (or any custom path) using Nginx reverse proxy.
+
+## Quick Start (Published Image - API Only)
+
+```bash
+docker-compose up
+```
+
+- **API:** http://localhost:8080/hindsight/docs
+- **Control Plane:** http://localhost:9999 (direct access, not proxied)
+
+## Full Stack with Custom Base Path (Requires Build)
+
+**Important:** You cannot rebuild from the published image with build args. You must build from source.
+
+### Build from Source with Custom Base Path
+
+1. **Clone the repository** (if you haven't):
+```bash
+git clone https://github.com/vectorize-io/hindsight.git
+cd hindsight
+```
+
+2. **Build with base path**:
+```bash
+docker build \
+  --build-arg NEXT_PUBLIC_BASE_PATH=/hindsight \
+  -f docker/standalone/Dockerfile \
+  -t hindsight:custom \
+  .
+```
+
+3. **Update docker-compose.yml** to use your built image:
+```yaml
+services:
+  hindsight:
+    image: hindsight:custom  # ← Change this
+    environment:
+      HINDSIGHT_API_BASE_PATH: /hindsight
+      NEXT_PUBLIC_BASE_PATH: /hindsight
+```
+
+4. **Update nginx.conf** to handle Control Plane routes (see below)
+
+5. **Run**:
+```bash
+docker-compose up
+```
+
+### Required nginx.conf for Full Stack
+
+Replace the current `nginx.conf` with this to proxy both API and Control Plane:
+
+```nginx
+events { worker_connections 1024; }
+
+http {
+    include /etc/nginx/mime.types;
+    default_type application/octet-stream;
+
+    upstream hindsight_api { server hindsight:8888; }
+    upstream hindsight_cp { server hindsight:9999; }
+
+    server {
+        listen 80;
+
+        # API
+        location ~ ^/hindsight/(docs|openapi\.json|health|metrics|v1|mcp) {
+            proxy_pass http://hindsight_api;
+            proxy_set_header Host $http_host;
+        }
+
+        # Control Plane static files
+        location ~ ^/hindsight/_next/ {
+            proxy_pass http://hindsight_cp;
+            proxy_set_header Host $http_host;
+        }
+
+        # Control Plane UI
+        location /hindsight {
+            proxy_pass http://hindsight_cp;
+            proxy_set_header Host $http_host;
+        }
+
+        location = / { return 301 /hindsight; }
+    }
+}
+```
+
+### Why Build is Required
+
+Next.js requires `basePath` at **build time**. The published image was built without a custom base path, so you must rebuild from source with the `NEXT_PUBLIC_BASE_PATH` build arg to deploy the Control Plane under a subpath.
+
+The API works without rebuild because `HINDSIGHT_API_BASE_PATH` is a runtime environment variable.

docker/docker-compose/nginx/docker-compose.yml

@@ -0,0 +1,88 @@
+# Hindsight API deployment with Nginx reverse proxy (API-only)
+#
+# This example deploys Hindsight API under the path /hindsight with:
+# - Hindsight standalone image (API + Control Plane + embedded pg0)
+# - Nginx reverse proxy (API only)
+#
+# Quick Start:
+#   docker-compose -f docker/docker-compose/nginx/docker-compose.yml up
+#
+# Access:
+#   API (via nginx):         http://localhost:8080/hindsight/docs
+#   Control Plane (direct):  http://localhost:9999
+#
+# For full stack deployment (API + Control Plane both under /hindsight):
+#   See README.md in this directory for instructions on building with basePath.
+#
+# Note: This configuration uses the published image (no build required).
+#       Control Plane is served directly because Next.js basePath requires
+#       build-time configuration. See README.md for the full stack option.
+
+services:
+  # Hindsight (API + Control Plane + embedded pg0)
+  hindsight:
+    image: ghcr.io/vectorize-io/hindsight:latest
+    ports:
+      - "9999:9999"  # Control Plane (direct access, not proxied)
+    environment:
+      # API base path for reverse proxy
+      HINDSIGHT_API_BASE_PATH: /hindsight
+
+      # LLM configuration
+      # Using mock provider for testing (no API key needed)
+      # For production, set OPENAI_API_KEY or ANTHROPIC_API_KEY and use a real provider
+      HINDSIGHT_API_LLM_PROVIDER: ${HINDSIGHT_API_LLM_PROVIDER:-mock}
+      HINDSIGHT_API_LLM_API_KEY: ${OPENAI_API_KEY:-not-needed-for-mock}
+      HINDSIGHT_API_LLM_MODEL: ${HINDSIGHT_API_LLM_MODEL:-mock-model}
+
+      # Production examples (uncomment and set appropriate API key):
+      # HINDSIGHT_API_LLM_PROVIDER: openai
+      # HINDSIGHT_API_LLM_API_KEY: ${OPENAI_API_KEY}
+      # HINDSIGHT_API_LLM_MODEL: gpt-4o-mini
+
+      # HINDSIGHT_API_LLM_PROVIDER: anthropic
+      # HINDSIGHT_API_LLM_API_KEY: ${ANTHROPIC_API_KEY}
+      # HINDSIGHT_API_LLM_MODEL: claude-sonnet-4-20250514
+
+      # Server config
+      HINDSIGHT_API_HOST: 0.0.0.0
+      HINDSIGHT_API_PORT: 8888
+      HINDSIGHT_API_LOG_LEVEL: info
+
+      # Control Plane config
+      HINDSIGHT_CP_DATAPLANE_API_URL: http://localhost:8888
+    volumes:
+      # Persist embedded pg0 database
+      - hindsight_data:/app/data
+    # Note: Ports not exposed - access via Nginx at localhost:8080/hindsight/
+    # To debug directly, uncomment these ports:
+    # ports:
+    #   - "8888:8888"  # API
+    #   - "9999:9999"  # Control Plane
+    healthcheck:
+      test: ["CMD", "curl", "-f", "http://localhost:8888/hindsight/health"]
+      interval: 10s
+      timeout: 5s
+      retries: 3
+      start_period: 30s
+    networks:
+      - hindsight
+
+  # Nginx reverse proxy
+  nginx:
+    image: nginx:alpine
+    ports:
+      - "8080:80"
+    volumes:
+      - ./nginx.conf:/etc/nginx/nginx.conf:ro
+    depends_on:
+      hindsight:
+        condition: service_healthy
+    networks:
+      - hindsight
+
+volumes:
+  hindsight_data:
+
+networks:
+  hindsight:

docker/docker-compose/nginx/nginx.conf

@@ -0,0 +1,40 @@
+# Nginx configuration for API-only reverse proxy
+# Control Plane accessed directly (not through nginx)
+
+events {
+    worker_connections 1024;
+}
+
+http {
+    include /etc/nginx/mime.types;
+    default_type application/octet-stream;
+
+    # Logging
+    access_log /var/log/nginx/access.log;
+    error_log /var/log/nginx/error.log;
+
+    # Upstream - Hindsight API
+    upstream hindsight_api {
+        server hindsight:8888;
+    }
+
+    server {
+        listen 80;
+        server_name _;
+
+        # API endpoints - forward with /hindsight prefix
+        location /hindsight/ {
+            proxy_pass http://hindsight_api;
+
+            proxy_set_header Host $http_host;
+            proxy_set_header X-Real-IP $remote_addr;
+            proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
+            proxy_set_header X-Forwarded-Proto $scheme;
+        }
+
+        # Redirect root to API docs
+        location = / {
+            return 301 /hindsight/docs;
+        }
+    }
+}

docker/docker-compose/pg_textsearch/Dockerfile

@@ -0,0 +1,32 @@
+# PostgreSQL with pgvector and pg_textsearch extensions
+# Note: pg_textsearch requires PostgreSQL 17+
+FROM postgres:17
+
+# Install build dependencies
+RUN apt-get update && apt-get install -y \
+    build-essential \
+    git \
+    postgresql-server-dev-17 \
+    libpq-dev \
+    && rm -rf /var/lib/apt/lists/*
+
+# Install pgvector
+RUN cd /tmp && \
+    git clone --branch v0.8.0 https://github.com/pgvector/pgvector.git && \
+    cd pgvector && \
+    make && \
+    make install
+
+# Install pg_textsearch
+RUN cd /tmp && \
+    git clone https://github.com/timescale/pg_textsearch.git && \
+    cd pg_textsearch && \
+    make && \
+    make install
+
+# Clean up source files and build dependencies
+RUN rm -rf /tmp/pgvector /tmp/pg_textsearch && \
+    apt-get purge -y --auto-remove build-essential git postgresql-server-dev-17
+
+# Ensure extensions are preloaded
+RUN echo "shared_preload_libraries = 'pg_textsearch'" >> /usr/share/postgresql/postgresql.conf.sample

docker/docker-compose/pg_textsearch/docker-compose.yaml

@@ -0,0 +1,91 @@
+name: hindsight
+# Docker Compose file for Hindsight with PostgreSQL and Timescale pg_textsearch
+# docker compose -f docker/docker-compose/pg_textsearch/docker-compose.yaml down && sleep 2 && docker compose -f docker/docker-compose/pg_textsearch/docker-compose.yaml up -d
+# Make sure to set the required environment variables before running:
+# - HINDSIGHT_DB_PASSWORD: Password for the PostgreSQL user
+# - Configure LLM provider variables as needed (see below in the hindsight service)
+#
+# Usage:
+# docker compose up -d
+#
+# Optional environment variables with defaults:
+# - HINDSIGHT_VERSION: Hindsight application version (default: latest)
+# - HINDSIGHT_DB_USER: PostgreSQL user (default: hindsight_user)
+# - HINDSIGHT_DB_NAME: PostgreSQL database name (default: hindsight_db)
+
+services:
+  db:
+    # Use custom PostgreSQL image with pgvector and pg_textsearch extensions
+    build:
+      context: .
+      dockerfile: Dockerfile
+    container_name: hindsight-db
+    restart: always
+    # Expose PostgreSQL port
+    ports:
+      - "5437:5432"
+    environment:
+      POSTGRES_USER: ${HINDSIGHT_DB_USER:-hindsight_user}
+      POSTGRES_PASSWORD: ${HINDSIGHT_DB_PASSWORD:-hindsight_password}
+      POSTGRES_DB: ${HINDSIGHT_DB_NAME:-hindsight_db}
+    volumes:
+      - pg_data:/var/lib/postgresql/data
+    networks:
+      - hindsight-net
+
+  pg-textsearch-init:
+    build:
+      context: .
+      dockerfile: Dockerfile
+    depends_on:
+      - db
+    environment:
+      - PGPASSWORD=${HINDSIGHT_DB_PASSWORD:-hindsight_password}
+    command: >
+      bash -c "
+        echo 'Waiting for PostgreSQL to be ready...';
+        until pg_isready -h hindsight-db -p 5432 -U hindsight_user; do
+          echo 'PostgreSQL is unavailable - sleeping';
+          sleep 2;
+        done;
+        echo 'PostgreSQL is ready - creating hindsight_db database';
+        psql -h hindsight-db -p 5432 -U hindsight_user -c 'CREATE DATABASE hindsight_db;' 2>/dev/null || echo 'Database already exists';
+        echo 'Creating extensions in hindsight_db database';
+        psql -h hindsight-db -p 5432 -U hindsight_user -d hindsight_db -c 'CREATE EXTENSION IF NOT EXISTS vector CASCADE;';
+        psql -h hindsight-db -p 5432 -U hindsight_user -d hindsight_db -c 'CREATE EXTENSION IF NOT EXISTS pg_textsearch CASCADE;';
+        echo 'Database and extensions created successfully';
+      "
+    restart: "no"
+    networks:
+      - hindsight-net
+
+  hindsight:
+    image: ghcr.io/vectorize-io/hindsight:${HINDSIGHT_VERSION:-latest}
+    container_name: hindsight-app
+    ports:
+      - "8888:8888"
+      - "9999:9999"
+    environment:
+      # LLM Configuration
+      HINDSIGHT_API_LLM_PROVIDER: ${HINDSIGHT_API_LLM_PROVIDER:-openai}
+      HINDSIGHT_API_LLM_API_KEY: ${OPENAI_API_KEY:-your-api-key}
+
+      # Database Configuration
+      HINDSIGHT_API_DATABASE_URL: postgresql://${HINDSIGHT_DB_USER:-hindsight_user}:${HINDSIGHT_DB_PASSWORD:-hindsight_password}@db:5432/${HINDSIGHT_DB_NAME:-hindsight_db}
+
+      # Vector and Text Search Extensions
+      HINDSIGHT_API_VECTOR_EXTENSION: pgvector
+      HINDSIGHT_API_TEXT_SEARCH_EXTENSION: pg_textsearch
+
+    depends_on:
+      - db
+    networks:
+      - hindsight-net
+
+
+networks:
+  hindsight-net:
+    driver: bridge
+
+volumes:
+  pg_data:

docker/docker-compose/s3-file-storage/docker-compose.yaml

@@ -0,0 +1,83 @@
+# Docker Compose file for Hindsight with S3 file storage (SeaweedFS)
+#
+# SeaweedFS (Apache 2.0) provides an S3-compatible object storage backend
+# for storing uploaded files instead of PostgreSQL BYTEA storage.
+#
+# Make sure to set the required environment variables before running:
+# - HINDSIGHT_DB_PASSWORD: Password for the PostgreSQL user
+# - Configure LLM provider variables as needed (see below in the hindsight service)
+#
+# Usage:
+# docker compose up -d
+#
+# Optional environment variables with defaults:
+# - HINDSIGHT_VERSION: Hindsight application version (default: latest)
+# - HINDSIGHT_DB_USER: PostgreSQL user (default: hindsight_user)
+# - HINDSIGHT_DB_NAME: PostgreSQL database name (default: hindsight_db)
+# - HINDSIGHT_DB_VERSION: PostgreSQL version (default: 18)
+# - SEAWEEDFS_S3_ACCESS_KEY: S3 access key (default: hindsight_s3_key)
+# - SEAWEEDFS_S3_SECRET_KEY: S3 secret key (default: hindsight_s3_secret)
+
+services:
+  db:
+    image: pgvector/pgvector:pg${HINDSIGHT_DB_VERSION:-18}
+    container_name: hindsight-db
+    restart: always
+    environment:
+      POSTGRES_USER: ${HINDSIGHT_DB_USER:-hindsight_user}
+      POSTGRES_PASSWORD: ${HINDSIGHT_DB_PASSWORD:?Please set the HINDSIGHT_DB_PASSWORD env variable}
+      POSTGRES_DB: ${HINDSIGHT_DB_NAME:-hindsight_db}
+    volumes:
+      - pg_data:/var/lib/postgresql/${HINDSIGHT_DB_VERSION:-18}/docker
+    networks:
+      - hindsight-net
+
+  seaweedfs:
+    image: chrislusf/seaweedfs:latest
+    container_name: hindsight-seaweedfs
+    restart: always
+    # Single-node mode: master + volume + filer + S3 gateway all in one process
+    command: >
+      server
+      -s3
+      -s3.port=8333
+      -s3.config=/etc/seaweedfs/s3.json
+      -ip.bind=0.0.0.0
+    volumes:
+      - seaweedfs_data:/data
+      - ./s3.json:/etc/seaweedfs/s3.json:ro
+    # Expose S3 API port (uncomment to access from host)
+    # ports:
+    #   - "8333:8333"
+    networks:
+      - hindsight-net
+
+  hindsight:
+    image: ghcr.io/vectorize-io/hindsight:${HINDSIGHT_VERSION:-latest}
+    container_name: hindsight-app
+    ports:
+      - "8888:8888"
+      - "9999:9999"
+    environment:
+      - HINDSIGHT_API_LLM_API_KEY=${OPENAI_API_KEY?Please set the OPENAI_API_KEY env variable}
+      - HINDSIGHT_API_DATABASE_URL=postgresql://${HINDSIGHT_DB_USER:-hindsight_user}:${HINDSIGHT_DB_PASSWORD:?Please set the HINDSIGHT_DB_PASSWORD env variable}@db:5432/${HINDSIGHT_DB_NAME:-hindsight_db}
+      # S3 file storage configuration (SeaweedFS)
+      - HINDSIGHT_API_FILE_STORAGE_TYPE=s3
+      - HINDSIGHT_API_FILE_STORAGE_S3_BUCKET=hindsight
+      - HINDSIGHT_API_FILE_STORAGE_S3_ENDPOINT=http://seaweedfs:8333
+      - HINDSIGHT_API_FILE_STORAGE_S3_REGION=us-east-1
+      - HINDSIGHT_API_FILE_STORAGE_S3_ACCESS_KEY_ID=${SEAWEEDFS_S3_ACCESS_KEY:-hindsight_s3_key}
+      - HINDSIGHT_API_FILE_STORAGE_S3_SECRET_ACCESS_KEY=${SEAWEEDFS_S3_SECRET_KEY:-hindsight_s3_secret}
+    depends_on:
+      - db
+      - seaweedfs
+    networks:
+      - hindsight-net
+
+networks:
+  hindsight-net:
+    driver: bridge
+
+volumes:
+  pg_data:
+  seaweedfs_data:

docker/docker-compose/s3-file-storage/s3.json

@@ -0,0 +1,19 @@
+{
+  "identities": [
+    {
+      "name": "hindsight",
+      "credentials": [
+        {
+          "accessKey": "hindsight_s3_key",
+          "secretKey": "hindsight_s3_secret"
+        }
+      ],
+      "actions": [
+        "Admin",
+        "Read",
+        "Write",
+        "List"
+      ]
+    }
+  ]
+}

docker/docker-compose/timescale/.dockerignore

@@ -0,0 +1,16 @@
+# Git
+.git
+.gitignore
+.gitattributes
+
+# Docker
+docker-compose.yaml
+.dockerignore
+
+# Documentation
+README.md
+*.md
+
+# Environment
+.env
+.env.example

docker/docker-compose/timescale/.env.example

@@ -0,0 +1,25 @@
+# PostgreSQL Configuration
+HINDSIGHT_DB_USER=hindsight_user
+HINDSIGHT_DB_PASSWORD=change-me-to-secure-password
+HINDSIGHT_DB_NAME=hindsight_db
+
+# Hindsight Version
+HINDSIGHT_VERSION=latest
+
+# LLM Configuration
+HINDSIGHT_API_LLM_PROVIDER=openai
+OPENAI_API_KEY=your-openai-api-key-here
+
+# Alternative LLM providers (uncomment and configure as needed):
+# HINDSIGHT_API_LLM_PROVIDER=anthropic
+# ANTHROPIC_API_KEY=your-anthropic-api-key
+
+# HINDSIGHT_API_LLM_PROVIDER=gemini
+# GEMINI_API_KEY=your-gemini-api-key
+
+# HINDSIGHT_API_LLM_PROVIDER=groq
+# GROQ_API_KEY=your-groq-api-key
+
+# Vector and Text Search (already configured in docker-compose.yaml)
+# HINDSIGHT_API_VECTOR_EXTENSION=pgvectorscale
+# HINDSIGHT_API_TEXT_SEARCH_EXTENSION=pg_textsearch

docker/docker-compose/timescale/Dockerfile

@@ -0,0 +1,55 @@
+# PostgreSQL with pgvector, pgvectorscale, and pg_textsearch extensions
+# All three extensions from Timescale/pgvector for high-performance vector and text search
+# Note: Requires PostgreSQL 16+
+FROM postgres:17
+
+# Install build dependencies and Rust toolchain
+RUN apt-get update && apt-get install -y \
+    build-essential \
+    git \
+    postgresql-server-dev-17 \
+    libpq-dev \
+    cmake \
+    curl \
+    pkg-config \
+    libssl-dev \
+    && rm -rf /var/lib/apt/lists/*
+
+# Install Rust toolchain (required for pgvectorscale)
+RUN curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y
+ENV PATH="/root/.cargo/bin:${PATH}"
+
+# Install pgvector (required by pgvectorscale)
+RUN cd /tmp && \
+    git clone --branch v0.8.0 https://github.com/pgvector/pgvector.git && \
+    cd pgvector && \
+    make && \
+    make install && \
+    rm -rf /tmp/pgvector
+
+# Install cargo-pgrx (PostgreSQL extension framework for Rust)
+RUN cargo install cargo-pgrx --version 0.12.5 --locked && \
+    cargo pgrx init --pg17 /usr/bin/pg_config
+
+# Install pgvectorscale (DiskANN index support)
+RUN cd /tmp && \
+    git clone --branch 0.5.1 https://github.com/timescale/pgvectorscale.git && \
+    cd pgvectorscale/pgvectorscale && \
+    cargo pgrx install --release && \
+    rm -rf /tmp/pgvectorscale
+
+# Install pg_textsearch (BM25 text search)
+RUN cd /tmp && \
+    git clone https://github.com/timescale/pg_textsearch.git && \
+    cd pg_textsearch && \
+    make && \
+    make install && \
+    rm -rf /tmp/pg_textsearch
+
+# Clean up build dependencies (keep runtime dependencies)
+RUN apt-get purge -y --auto-remove git cmake curl && \
+    rm -rf /root/.cargo/registry /root/.cargo/git
+
+# Ensure extensions are preloaded (pg_textsearch requires preloading)
+RUN echo "shared_preload_libraries = 'pg_textsearch'" >> /usr/share/postgresql/postgresql.conf.sample
+

docker/docker-compose/timescale/README.md

@@ -0,0 +1,101 @@
+# Hindsight with Timescale Extensions
+
+This Docker Compose setup provides a complete Hindsight deployment with **Timescale extensions**:
+- **pgvectorscale** - DiskANN algorithm for disk-based scalable vector search  
+- **pg_textsearch** - High-performance BM25 text search
+
+Both extensions are from [Timescale](https://github.com/timescale) and provide production-grade performance.
+
+## Prerequisites
+
+- Docker and Docker Compose installed
+- OpenAI API key (or another LLM provider)
+
+## Quick Start
+
+```bash
+# Set environment variables
+export HINDSIGHT_DB_PASSWORD="your-secure-password"
+export OPENAI_API_KEY="your-openai-api-key"
+
+# Build and start
+docker compose -f docker/docker-compose/timescale/docker-compose.yaml up -d --build
+
+# Check logs
+
+docker compose -f docker/docker-compose/timescale/docker-compose.yaml logs -f
+```
+
+**Access:**
+- API: http://localhost:8888
+- Control Plane: http://localhost:9999
+
+## Stop and Clean Up
+
+```bash
+# Stop services
+docker compose -f docker/docker-compose/timescale/docker-compose.yaml down
+
+# Remove volumes (deletes all data)
+docker compose -f docker/docker-compose/timescale/docker-compose.yaml down -v
+```
+
+## Configuration
+
+### Environment Variables
+
+| Variable | Description | Default |
+|----------|-------------|---------|
+| `HINDSIGHT_DB_PASSWORD` | PostgreSQL password | `hindsight_password` |
+| `HINDSIGHT_DB_USER` | PostgreSQL username | `hindsight_user` |
+| `HINDSIGHT_DB_NAME` | Database name | `hindsight_db` |
+| `HINDSIGHT_VERSION` | Hindsight Docker image version | `latest` |
+| `OPENAI_API_KEY` | OpenAI API key | (required) |
+| `HINDSIGHT_API_LLM_PROVIDER` | LLM provider | `openai` |
+
+### Why Timescale Extensions?
+
+**pgvectorscale (DiskANN):**
+- 28x lower p95 latency vs dedicated vector databases
+- 16x higher query throughput at 99% recall
+- 60-75% cost reduction (disk is cheaper than RAM)
+- Best for large datasets (10M+ vectors)
+
+**pg_textsearch (BM25):**
+- High-performance keyword retrieval
+- Native BM25 ranking algorithm
+- Optimized for full-text search
+
+## Troubleshooting
+
+### Extensions not installed
+
+Check if extensions are available:
+
+```bash
+docker exec -it hindsight-db-timescale psql -U hindsight_user -d hindsight_db -c "\dx"
+```
+
+You should see:
+- `vector` (pgvector)
+- `vectorscale` (pgvectorscale/DiskANN)
+- `pg_textsearch` (BM25 search)
+
+### Build fails
+
+If the Docker build fails during pgvectorscale compilation:
+
+1. Ensure you have sufficient memory (recommended: 4GB+)
+2. Check Docker build logs for Rust compilation errors
+3. Try building with more resources: `docker compose build --no-cache --memory 4g`
+
+### Port conflicts
+
+If port 5438 is already in use, modify the `ports` section in docker-compose.yaml.
+
+## Learn More
+
+- [pgvectorscale GitHub](https://github.com/timescale/pgvectorscale)
+- [pg_textsearch GitHub](https://github.com/timescale/pg_textsearch)
+- [HNSW vs DiskANN](https://www.tigerdata.com/learn/hnsw-vs-diskann)
+- [Hindsight Documentation](https://hindsight.dev)

docker/docker-compose/timescale/docker-compose.yaml

@@ -0,0 +1,108 @@
+name: hindsight
+# Docker Compose file for Hindsight with Timescale extensions
+# - pgvectorscale: DiskANN vector search (disk-based, scalable)
+# - pg_textsearch: BM25 text search (high-performance keyword retrieval)
+#
+# Quick start:
+# docker compose -f docker/docker-compose/timescale/docker-compose.yaml up -d --build
+#
+# Required environment variables:
+# - HINDSIGHT_DB_PASSWORD: Password for the PostgreSQL user
+# - OPENAI_API_KEY (or configure another LLM provider)
+#
+# Optional environment variables with defaults:
+# - HINDSIGHT_VERSION: Hindsight application version (default: latest)
+# - HINDSIGHT_DB_USER: PostgreSQL user (default: hindsight_user)
+# - HINDSIGHT_DB_NAME: PostgreSQL database name (default: hindsight_db)
+
+services:
+  db:
+    # Custom PostgreSQL image with Timescale extensions (pgvectorscale + pg_textsearch)
+    build:
+      context: .
+      dockerfile: Dockerfile
+    container_name: hindsight-db-timescale
+    restart: always
+    # Expose PostgreSQL port (using 5438 to avoid conflicts with other setups)
+    ports:
+      - "5438:5432"
+    environment:
+      POSTGRES_USER: ${HINDSIGHT_DB_USER:-hindsight_user}
+      POSTGRES_PASSWORD: ${HINDSIGHT_DB_PASSWORD:-hindsight_password}
+      POSTGRES_DB: ${HINDSIGHT_DB_NAME:-hindsight_db}
+    volumes:
+      - pg_data:/var/lib/postgresql/data
+    networks:
+      - hindsight-net
+    # Health check to ensure database is ready
+    healthcheck:
+      test: ["CMD-SHELL", "pg_isready -U hindsight_user"]
+      interval: 5s
+      timeout: 5s
+      retries: 5
+
+  timescale-init:
+    build:
+      context: .
+      dockerfile: Dockerfile
+    depends_on:
+      db:
+        condition: service_healthy
+    environment:
+      - PGPASSWORD=${HINDSIGHT_DB_PASSWORD:-hindsight_password}
+    command: >
+      bash -c "
+        echo 'PostgreSQL is ready - creating hindsight_db database';
+        psql -h hindsight-db-timescale -p 5432 -U hindsight_user -c 'CREATE DATABASE hindsight_db;' 2>/dev/null || echo 'Database already exists';
+        echo 'Installing Timescale extensions...';
+        echo '1/3: Installing pgvector (required by pgvectorscale)...';
+        psql -h hindsight-db-timescale -p 5432 -U hindsight_user -d hindsight_db -c 'CREATE EXTENSION IF NOT EXISTS vector CASCADE;';
+        echo '2/3: Installing pgvectorscale (DiskANN vector search)...';
+        psql -h hindsight-db-timescale -p 5432 -U hindsight_user -d hindsight_db -c 'CREATE EXTENSION IF NOT EXISTS vectorscale CASCADE;';
+        echo '3/3: Installing pg_textsearch (BM25 text search)...';
+        psql -h hindsight-db-timescale -p 5432 -U hindsight_user -d hindsight_db -c 'CREATE EXTENSION IF NOT EXISTS pg_textsearch CASCADE;';
+        echo '';
+        echo '✅ Timescale extensions installed successfully';
+        echo '';
+        echo 'Installed extensions:';
+        psql -h hindsight-db-timescale -p 5432 -U hindsight_user -d hindsight_db -c \"\\dx\" | grep -E '(vector|vectorscale|pg_textsearch)';
+      "
+    restart: "no"
+    networks:
+      - hindsight-net
+
+  hindsight:
+    image: ghcr.io/vectorize-io/hindsight:${HINDSIGHT_VERSION:-latest}
+    container_name: hindsight-app-timescale
+    ports:
+      - "8888:8888"
+      - "9999:9999"
+    environment:
+      # LLM Configuration
+      HINDSIGHT_API_LLM_PROVIDER: ${HINDSIGHT_API_LLM_PROVIDER:-openai}
+      HINDSIGHT_API_LLM_API_KEY: ${OPENAI_API_KEY:-your-api-key}
+
+      # Database Configuration
+      HINDSIGHT_API_DATABASE_URL: postgresql://${HINDSIGHT_DB_USER:-hindsight_user}:${HINDSIGHT_DB_PASSWORD:-hindsight_password}@db:5432/${HINDSIGHT_DB_NAME:-hindsight_db}
+
+      # Timescale Extensions
+      # pgvectorscale: DiskANN algorithm for disk-based scalable vector search
+      HINDSIGHT_API_VECTOR_EXTENSION: pgvectorscale
+      # pg_textsearch: High-performance BM25 text search
+      HINDSIGHT_API_TEXT_SEARCH_EXTENSION: pg_textsearch
+
+    depends_on:
+      db:
+        condition: service_healthy
+      timescale-init:
+        condition: service_completed_successfully
+    networks:
+      - hindsight-net
+
+
+networks:
+  hindsight-net:
+    driver: bridge
+
+volumes:
+  pg_data:

docker/docker-compose/vchord/docker-compose.yaml

@@ -0,0 +1,93 @@
+name: hindsight
+# Docker Compose file for Hindsight with PostgreSQL and vectorchord
+# docker compose -f docker/docker-compose/docker-compose.yaml down && sleep 2 && docker compose -f docker/docker-compose/docker-compose.yaml up -d
+# Make sure to set the required environment variables before running:
+# - HINDSIGHT_DB_PASSWORD: Password for the PostgreSQL user
+# - Configure LLM provider variables as needed (see below in the hindsight service)
+#
+# Usage:
+# docker compose up -d
+#
+# Optional environment variables with defaults:
+# - HINDSIGHT_VERSION: Hindsight application version (default: latest)
+# - HINDSIGHT_DB_USER: PostgreSQL user (default: hindsight_user)
+# - HINDSIGHT_DB_NAME: PostgreSQL database name (default: hindsight_db)
+# - HINDSIGHT_DB_VERSION: PostgreSQL version (default: 18)
+
+services:
+  db:
+    # Use a PostgreSQL-Image with vectorchord extension pre-installed
+    image: tensorchord/vchord-suite:pg${HINDSIGHT_DB_VERSION:-18-latest}
+    container_name: hindsight-db
+    restart: always
+    # Expose PostgreSQL port
+    ports:
+      - "5436:5432"
+    environment:
+      POSTGRES_USER: ${HINDSIGHT_DB_USER:-hindsight_user}
+      POSTGRES_PASSWORD: ${HINDSIGHT_DB_PASSWORD:-hindsight_password}
+      POSTGRES_DB: ${HINDSIGHT_DB_NAME:-hindsight_db}
+    volumes:
+      - pg_data:/var/lib/postgresql/${HINDSIGHT_DB_VERSION:-18}/docker
+    networks:
+      - hindsight-net
+
+  vectorchord-init:
+    image: tensorchord/vchord-suite:pg18-latest
+    #container_name: vectorchord-init
+    depends_on:
+      - db
+    environment:
+      - PGPASSWORD=${HINDSIGHT_DB_PASSWORD:-hindsight_password}
+    command: >
+      bash -c "
+        echo 'Waiting for PostgreSQL to be ready...';
+        until pg_isready -h hindsight-db -p 5432 -U hindsight_user; do
+          echo 'PostgreSQL is unavailable - sleeping';
+          sleep 2;
+        done;
+        echo 'PostgreSQL is ready - creating hindsight_db database';
+        psql -h hindsight-db -p 5432 -U hindsight_user -c 'CREATE DATABASE hindsight_db;' 2>/dev/null || echo 'Database already exists';
+        echo 'Creating extensions in hindsight_db database';
+        psql -h hindsight-db -p 5432 -U hindsight_user -d hindsight_db -c 'CREATE EXTENSION IF NOT EXISTS vchord CASCADE;';
+        psql -h hindsight-db -p 5432 -U hindsight_user -d hindsight_db -c 'CREATE EXTENSION IF NOT EXISTS pg_tokenizer CASCADE;';
+        psql -h hindsight-db -p 5432 -U hindsight_user -d hindsight_db -c 'CREATE EXTENSION IF NOT EXISTS vchord_bm25 CASCADE;';
+        echo 'Creating llmlingua2 tokenizer';
+        psql -h hindsight-db -p 5432 -U hindsight_user -d hindsight_db -c \"SELECT create_tokenizer('llmlingua2', \\$\\$ model = \\\"llmlingua2\\\" \\$\\$);\" 2>/dev/null || echo 'Tokenizer already exists or creation skipped';
+        echo 'Database and extensions created successfully';
+      "
+    restart: "no"
+    networks:
+      - hindsight-net
+
+  hindsight:
+    image: ghcr.io/vectorize-io/hindsight:${HINDSIGHT_VERSION:-latest}
+    container_name: hindsight-app
+    ports:
+      - "8888:8888"
+      - "9999:9999"
+    environment:
+      # LLM Configuration (uses OpenAI for testing vchord)
+      # LLM configuration
+      HINDSIGHT_API_LLM_PROVIDER: ${HINDSIGHT_API_LLM_PROVIDER:-openai}
+      HINDSIGHT_API_LLM_API_KEY: ${OPENAI_API_KEY:-your-api-key}
+
+      # Database Configuration
+      HINDSIGHT_API_DATABASE_URL: postgresql://${HINDSIGHT_DB_USER:-hindsight_user}:${HINDSIGHT_DB_PASSWORD:-hindsight_password}@db:5432/${HINDSIGHT_DB_NAME:-hindsight_db}
+
+      # Vector and Text Search Extensions
+      HINDSIGHT_API_VECTOR_EXTENSION: vchord
+      HINDSIGHT_API_TEXT_SEARCH_EXTENSION: vchord
+
+    depends_on:
+      - db
+    networks:
+      - hindsight-net
+
+
+networks:
+  hindsight-net:
+    driver: bridge
+
+volumes:
+  pg_data:

docker/standalone/Dockerfile

@@ -8,6 +8,7 @@
 #                                     Set to false when using external providers (TEI, OpenAI, Cohere)
 #   PRELOAD_ML_MODELS=true/false   - Pre-download ML models during build (default: true)
 #                                     Only effective when INCLUDE_LOCAL_MODELS=true
+#                                     NOTE: tiktoken encodings are ALWAYS preloaded (required for air-gapped deployments)
 #
 # Examples:
 #   docker build -t hindsight .                                          # Both (standalone)
@@ -111,6 +112,10 @@ RUN rm -f package-lock.json && sed -i '/"@vectorize-io\/hindsight-client":/d' pa
 # Copy built SDK directly into node_modules (more reliable than npm link in Docker)
 COPY --from=sdk-builder /app/hindsight-clients/typescript ./node_modules/@vectorize-io/hindsight-client
 
+# Accept base path as build argument for reverse proxy deployments
+# Usage: docker build --build-arg NEXT_PUBLIC_BASE_PATH=/hindsight ...
+ARG NEXT_PUBLIC_BASE_PATH=""
+
 # Build Control Plane - run next build first, then custom standalone copy
 # (The build:standalone script expects a specific path structure that differs in Docker)
 RUN npm exec -- next build
@@ -167,18 +172,57 @@ USER hindsight
 
 ENV PATH="/app/api/.venv/bin:${PATH}"
 
+# Pre-download tiktoken encoding (ALWAYS - required for token counting even in air-gapped envs)
+# Tiktoken is a core runtime dependency, not an optional ML model
+RUN MAX_RETRIES=3; \
+    RETRY_DELAY=5; \
+    for i in $(seq 1 $MAX_RETRIES); do \
+      echo "Attempt $i/$MAX_RETRIES: Downloading tiktoken encoding..."; \
+      /app/api/.venv/bin/python -c "\
+import tiktoken; \
+print('Downloading cl100k_base encoding...'); \
+tiktoken.get_encoding('cl100k_base'); \
+print('Tiktoken encoding cached successfully')" && break; \
+      if [ $i -lt $MAX_RETRIES ]; then \
+        echo "Attempt $i failed, retrying in ${RETRY_DELAY}s..."; \
+        sleep $RETRY_DELAY; \
+        RETRY_DELAY=$((RETRY_DELAY * 2)); \
+      fi; \
+    done; \
+    if [ $i -eq $MAX_RETRIES ]; then \
+      echo "ERROR: Failed to download tiktoken encoding after $MAX_RETRIES attempts"; \
+      exit 1; \
+    fi
+
 # Pre-download ML models to avoid runtime download (conditional)
 # Only runs if both PRELOAD_ML_MODELS=true AND INCLUDE_LOCAL_MODELS=true
+# Includes retry logic with exponential backoff for transient network failures
 ARG PRELOAD_ML_MODELS
 ARG INCLUDE_LOCAL_MODELS
+ENV HF_HUB_DOWNLOAD_TIMEOUT=600
 RUN if [ "$PRELOAD_ML_MODELS" = "true" ] && [ "$INCLUDE_LOCAL_MODELS" = "true" ]; then \
-    /app/api/.venv/bin/python -c "\
+    MAX_RETRIES=3; \
+    RETRY_DELAY=10; \
+    for i in $(seq 1 $MAX_RETRIES); do \
+      echo "Attempt $i/$MAX_RETRIES: Downloading ML models..."; \
+      /app/api/.venv/bin/python -c "\
+import os; os.environ['HF_HUB_DOWNLOAD_TIMEOUT'] = '600'; \
 from sentence_transformers import SentenceTransformer, CrossEncoder; \
 print('Downloading embedding model...'); \
 SentenceTransformer('BAAI/bge-small-en-v1.5'); \
 print('Downloading cross-encoder model...'); \
 CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2'); \
-print('Models cached successfully')"; \
+print('Models cached successfully')" && break; \
+      if [ $i -lt $MAX_RETRIES ]; then \
+        echo "Attempt $i failed, retrying in ${RETRY_DELAY}s..."; \
+        sleep $RETRY_DELAY; \
+        RETRY_DELAY=$((RETRY_DELAY * 2)); \
+      fi; \
+    done; \
+    if [ $i -eq $MAX_RETRIES ] && ! /app/api/.venv/bin/python -c "from sentence_transformers import SentenceTransformer; SentenceTransformer('BAAI/bge-small-en-v1.5')" 2>/dev/null; then \
+      echo "ERROR: Failed to download models after $MAX_RETRIES attempts"; \
+      exit 1; \
+    fi; \
     elif [ "$INCLUDE_LOCAL_MODELS" != "true" ]; then echo "Skipping ML model preload (local-models not included)"; \
     else echo "Skipping ML model preload"; fi
 
@@ -190,6 +234,10 @@ ENV HINDSIGHT_API_LOG_LEVEL=info
 ENV HINDSIGHT_ENABLE_API=true
 ENV HINDSIGHT_ENABLE_CP=false
 ENV PYTHONUNBUFFERED=1
+# Suppress verbose transformers/HuggingFace model loading warnings
+ENV TRANSFORMERS_VERBOSITY=error
+ENV HF_HUB_VERBOSITY=error
+ENV TOKENIZERS_PARALLELISM=false
 
 CMD ["/app/start-all.sh"]
 
@@ -275,18 +323,57 @@ USER hindsight
 
 ENV PATH="/app/api/.venv/bin:${PATH}"
 
+# Pre-download tiktoken encoding (ALWAYS - required for token counting even in air-gapped envs)
+# Tiktoken is a core runtime dependency, not an optional ML model
+RUN MAX_RETRIES=3; \
+    RETRY_DELAY=5; \
+    for i in $(seq 1 $MAX_RETRIES); do \
+      echo "Attempt $i/$MAX_RETRIES: Downloading tiktoken encoding..."; \
+      /app/api/.venv/bin/python -c "\
+import tiktoken; \
+print('Downloading cl100k_base encoding...'); \
+tiktoken.get_encoding('cl100k_base'); \
+print('Tiktoken encoding cached successfully')" && break; \
+      if [ $i -lt $MAX_RETRIES ]; then \
+        echo "Attempt $i failed, retrying in ${RETRY_DELAY}s..."; \
+        sleep $RETRY_DELAY; \
+        RETRY_DELAY=$((RETRY_DELAY * 2)); \
+      fi; \
+    done; \
+    if [ $i -eq $MAX_RETRIES ]; then \
+      echo "ERROR: Failed to download tiktoken encoding after $MAX_RETRIES attempts"; \
+      exit 1; \
+    fi
+
 # Pre-download ML models to avoid runtime download (conditional)
 # Only runs if both PRELOAD_ML_MODELS=true AND INCLUDE_LOCAL_MODELS=true
+# Includes retry logic with exponential backoff for transient network failures
 ARG PRELOAD_ML_MODELS
 ARG INCLUDE_LOCAL_MODELS
+ENV HF_HUB_DOWNLOAD_TIMEOUT=600
 RUN if [ "$PRELOAD_ML_MODELS" = "true" ] && [ "$INCLUDE_LOCAL_MODELS" = "true" ]; then \
-    /app/api/.venv/bin/python -c "\
+    MAX_RETRIES=3; \
+    RETRY_DELAY=10; \
+    for i in $(seq 1 $MAX_RETRIES); do \
+      echo "Attempt $i/$MAX_RETRIES: Downloading ML models..."; \
+      /app/api/.venv/bin/python -c "\
+import os; os.environ['HF_HUB_DOWNLOAD_TIMEOUT'] = '600'; \
 from sentence_transformers import SentenceTransformer, CrossEncoder; \
 print('Downloading embedding model...'); \
 SentenceTransformer('BAAI/bge-small-en-v1.5'); \
 print('Downloading cross-encoder model...'); \
 CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2'); \
-print('Models cached successfully')"; \
+print('Models cached successfully')" && break; \
+      if [ $i -lt $MAX_RETRIES ]; then \
+        echo "Attempt $i failed, retrying in ${RETRY_DELAY}s..."; \
+        sleep $RETRY_DELAY; \
+        RETRY_DELAY=$((RETRY_DELAY * 2)); \
+      fi; \
+    done; \
+    if [ $i -eq $MAX_RETRIES ] && ! /app/api/.venv/bin/python -c "from sentence_transformers import SentenceTransformer; SentenceTransformer('BAAI/bge-small-en-v1.5')" 2>/dev/null; then \
+      echo "ERROR: Failed to download models after $MAX_RETRIES attempts"; \
+      exit 1; \
+    fi; \
     elif [ "$INCLUDE_LOCAL_MODELS" != "true" ]; then echo "Skipping ML model preload (local-models not included)"; \
     else echo "Skipping ML model preload"; fi
 
@@ -300,6 +387,10 @@ ENV HINDSIGHT_CP_DATAPLANE_API_URL=http://localhost:8888
 ENV HINDSIGHT_ENABLE_API=true
 ENV HINDSIGHT_ENABLE_CP=true
 ENV PYTHONUNBUFFERED=1
+# Suppress verbose transformers/HuggingFace model loading warnings
+ENV TRANSFORMERS_VERBOSITY=error
+ENV HF_HUB_VERBOSITY=error
+ENV TOKENIZERS_PARALLELISM=false
 
 CMD ["/app/start-all.sh"]
 

docker/test-image.sh

@@ -6,28 +6,40 @@
 # Can be run locally or in CI pipelines.
 #
 # Usage:
-#   ./scripts/docker-smoke-test.sh <image> [target]
+#   ./docker/test-image.sh <image> [target]
 #
 # Arguments:
 #   image   - Docker image to test (e.g., hindsight-api:test, ghcr.io/vectorize-io/hindsight:latest)
 #   target  - Optional: 'cp-only' for control plane, otherwise assumes API image (default: api)
 #
 # Environment variables:
-#   GROQ_API_KEY              - Required for API/standalone images (LLM verification)
-#   HINDSIGHT_API_LLM_PROVIDER - LLM provider (default: groq)
-#   HINDSIGHT_API_LLM_MODEL    - LLM model (default: llama-3.3-70b-versatile)
-#   SMOKE_TEST_TIMEOUT        - Timeout in seconds (default: 120)
-#   SMOKE_TEST_CONTAINER_NAME - Container name (default: hindsight-smoke-test)
+#   GROQ_API_KEY                                - Required for API/standalone images (LLM verification)
+#   HINDSIGHT_API_LLM_PROVIDER                  - LLM provider (default: groq)
+#   HINDSIGHT_API_LLM_MODEL                     - LLM model (default: llama-3.3-70b-versatile)
+#   HINDSIGHT_API_EMBEDDINGS_PROVIDER           - Embeddings provider (optional, for slim images: openai, cohere, tei)
+#   HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY     - OpenAI API key for embeddings (optional)
+#   HINDSIGHT_API_RERANKER_PROVIDER             - Reranker provider (optional, for slim images: cohere, tei)
+#   HINDSIGHT_API_COHERE_API_KEY                - Cohere API key for reranking (optional)
+#   SMOKE_TEST_TIMEOUT                          - Timeout in seconds (default: 120)
+#   SMOKE_TEST_CONTAINER_NAME                   - Container name (default: hindsight-smoke-test)
 #
 # Examples:
-#   # Test a locally built image
-#   ./scripts/docker-smoke-test.sh hindsight-api:test
+#   # Test a locally built full image
+#   ./docker/test-image.sh hindsight-api:test
 #
 #   # Test a released image
-#   ./scripts/docker-smoke-test.sh ghcr.io/vectorize-io/hindsight:latest
+#   ./docker/test-image.sh ghcr.io/vectorize-io/hindsight:latest
 #
 #   # Test control plane image
-#   ./scripts/docker-smoke-test.sh hindsight-control-plane:test cp-only
+#   ./docker/test-image.sh hindsight-control-plane:test cp-only
+#
+#   # Test slim image with external providers
+#   export GROQ_API_KEY=gsk_xxx
+#   export HINDSIGHT_API_EMBEDDINGS_PROVIDER=openai
+#   export HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY=sk-xxx
+#   export HINDSIGHT_API_RERANKER_PROVIDER=cohere
+#   export HINDSIGHT_API_COHERE_API_KEY=xxx
+#   ./docker/test-image.sh hindsight-slim:test
 #
 # Exit codes:
 #   0 - Success (container healthy)
@@ -108,12 +120,32 @@ if [ "$TARGET" = "cp-only" ]; then
         -p "${HEALTH_PORT}:${HEALTH_PORT}" \
         "$IMAGE"
 else
-    docker run -d --name "$CONTAINER_NAME" \
-        -e HINDSIGHT_API_LLM_PROVIDER="$LLM_PROVIDER" \
-        -e HINDSIGHT_API_LLM_API_KEY="${GROQ_API_KEY}" \
-        -e HINDSIGHT_API_LLM_MODEL="$LLM_MODEL" \
-        -p "${HEALTH_PORT}:${HEALTH_PORT}" \
-        "$IMAGE"
+    # Build docker run command with required and optional env vars
+    DOCKER_CMD="docker run -d --name $CONTAINER_NAME"
+    DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_LLM_PROVIDER=$LLM_PROVIDER"
+    DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_LLM_API_KEY=${GROQ_API_KEY}"
+    DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_LLM_MODEL=$LLM_MODEL"
+
+    # Add optional embeddings provider config
+    if [ -n "${HINDSIGHT_API_EMBEDDINGS_PROVIDER:-}" ]; then
+        DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_EMBEDDINGS_PROVIDER=${HINDSIGHT_API_EMBEDDINGS_PROVIDER}"
+    fi
+    if [ -n "${HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY:-}" ]; then
+        DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY=${HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY}"
+    fi
+
+    # Add optional reranker provider config
+    if [ -n "${HINDSIGHT_API_RERANKER_PROVIDER:-}" ]; then
+        DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_RERANKER_PROVIDER=${HINDSIGHT_API_RERANKER_PROVIDER}"
+    fi
+    if [ -n "${HINDSIGHT_API_COHERE_API_KEY:-}" ]; then
+        DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_COHERE_API_KEY=${HINDSIGHT_API_COHERE_API_KEY}"
+    fi
+
+    DOCKER_CMD="$DOCKER_CMD -p ${HEALTH_PORT}:${HEALTH_PORT}"
+    DOCKER_CMD="$DOCKER_CMD $IMAGE"
+
+    eval $DOCKER_CMD
 fi
 
 # Wait for health endpoint

docker/test-slim-local.sh

@@ -0,0 +1,51 @@
+#!/bin/bash
+#
+# Local Test Script for Slim Docker Images
+#
+# This script makes it easy to test slim images locally with external providers.
+# It expects API keys to be set in environment variables.
+#
+# Usage:
+#   export GROQ_API_KEY=gsk_xxx
+#   export OPENAI_API_KEY=sk-xxx
+#   export COHERE_API_KEY=xxx
+#   ./docker/test-slim-local.sh
+#
+# Or inline:
+#   GROQ_API_KEY=gsk_xxx OPENAI_API_KEY=sk_xxx COHERE_API_KEY=xxx ./docker/test-slim-local.sh
+#
+
+set -euo pipefail
+
+# Check for required API keys
+if [ -z "${GROQ_API_KEY:-}" ]; then
+    echo "❌ Error: GROQ_API_KEY environment variable is required"
+    echo "Set it with: export GROQ_API_KEY=gsk_xxx"
+    exit 1
+fi
+
+if [ -z "${OPENAI_API_KEY:-}" ]; then
+    echo "❌ Error: OPENAI_API_KEY environment variable is required"
+    echo "Set it with: export OPENAI_API_KEY=sk-xxx"
+    exit 1
+fi
+
+if [ -z "${COHERE_API_KEY:-}" ]; then
+    echo "❌ Error: COHERE_API_KEY environment variable is required"
+    echo "Set it with: export COHERE_API_KEY=xxx"
+    exit 1
+fi
+
+# Configuration
+IMAGE="${1:-hindsight-slim:test}"
+echo "Testing image: $IMAGE"
+echo ""
+
+# Set up external providers
+export HINDSIGHT_API_EMBEDDINGS_PROVIDER=openai
+export HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY=$OPENAI_API_KEY
+export HINDSIGHT_API_RERANKER_PROVIDER=cohere
+export HINDSIGHT_API_COHERE_API_KEY=$COHERE_API_KEY
+
+# Run the test
+exec "$(dirname "$0")/test-image.sh" "$IMAGE" standalone

helm/hindsight/Chart.yaml

@@ -2,8 +2,8 @@ apiVersion: v2
 name: hindsight
 description: Hindsight helm chart
 type: application
-version: 0.2.1
-appVersion: "0.2.1"
+version: 0.4.11
+appVersion: "0.4.11"
 keywords:
   - ai
   - memory

helm/hindsight/templates/_helpers.tpl

@@ -80,6 +80,22 @@ Control plane selector labels
 app.kubernetes.io/component: control-plane
 {{- end }}
 
+{{/*
+Worker labels
+*/}}
+{{- define "hindsight.worker.labels" -}}
+{{ include "hindsight.labels" . }}
+app.kubernetes.io/component: worker
+{{- end }}
+
+{{/*
+Worker selector labels
+*/}}
+{{- define "hindsight.worker.selectorLabels" -}}
+{{ include "hindsight.selectorLabels" . }}
+app.kubernetes.io/component: worker
+{{- end }}
+
 {{/*
 Create the name of the service account to use
 */}}
@@ -111,6 +127,38 @@ API URL for control plane
 {{- printf "http://%s-api:%d" (include "hindsight.fullname" .) (.Values.api.service.port | int) }}
 {{- end }}
 
+{{/*
+TEI reranker labels
+*/}}
+{{- define "hindsight.tei.reranker.labels" -}}
+{{ include "hindsight.labels" . }}
+app.kubernetes.io/component: tei-reranker
+{{- end }}
+
+{{/*
+TEI reranker selector labels
+*/}}
+{{- define "hindsight.tei.reranker.selectorLabels" -}}
+{{ include "hindsight.selectorLabels" . }}
+app.kubernetes.io/component: tei-reranker
+{{- end }}
+
+{{/*
+TEI embedding labels
+*/}}
+{{- define "hindsight.tei.embedding.labels" -}}
+{{ include "hindsight.labels" . }}
+app.kubernetes.io/component: tei-embedding
+{{- end }}
+
+{{/*
+TEI embedding selector labels
+*/}}
+{{- define "hindsight.tei.embedding.selectorLabels" -}}
+{{ include "hindsight.selectorLabels" . }}
+app.kubernetes.io/component: tei-embedding
+{{- end }}
+
 {{/*
 Get the name of the secret to use
 */}}

helm/hindsight/templates/api-deployment.yaml

@@ -33,7 +33,7 @@ spec:
       - name: api
         securityContext:
           {{- toYaml .Values.securityContext | nindent 10 }}
-        image: "{{ .Values.api.image.repository }}:{{ .Values.api.image.tag | default .Values.version }}"
+        image: "{{ .Values.api.image.repository }}:{{ .Values.api.image.tag | default .Values.version | default .Chart.AppVersion }}"
         imagePullPolicy: {{ .Values.api.image.pullPolicy }}
         ports:
         - name: http
@@ -55,10 +55,30 @@ spec:
         {{- end }}
         - name: HINDSIGHT_API_DATABASE_URL
           value: {{ include "hindsight.databaseUrl" . | quote }}
+        {{- /* Disable internal worker when dedicated workers are enabled */}}
+        {{- if .Values.worker.enabled }}
+        - name: HINDSIGHT_API_WORKER_ENABLED
+          value: "false"
+        {{- end }}
+        {{- /* Explicitly set port to override K8s service discovery env var (HINDSIGHT_API_PORT) */}}
+        - name: HINDSIGHT_API_PORT
+          value: {{ .Values.api.service.targetPort | quote }}
         {{- range $key, $value := .Values.api.env }}
         - name: {{ $key }}
           value: {{ $value | quote }}
         {{- end }}
+        {{- if .Values.tei.reranker.enabled }}
+        - name: HINDSIGHT_API_RERANKER_PROVIDER
+          value: "tei"
+        - name: HINDSIGHT_API_RERANKER_TEI_URL
+          value: "http://{{ include "hindsight.fullname" . }}-tei-reranker:{{ .Values.tei.reranker.port }}"
+        {{- end }}
+        {{- if .Values.tei.embedding.enabled }}
+        - name: HINDSIGHT_API_EMBEDDINGS_PROVIDER
+          value: "tei"
+        - name: HINDSIGHT_API_EMBEDDINGS_TEI_URL
+          value: "http://{{ include "hindsight.fullname" . }}-tei-embedding:{{ .Values.tei.embedding.port }}"
+        {{- end }}
         {{- /* Only use api.secrets when not using existingSecret (for chart-managed secrets) */}}
         {{- if not .Values.existingSecret }}
         {{- range $key, $value := .Values.api.secrets }}
@@ -79,7 +99,7 @@ spec:
       nodeSelector:
         {{- toYaml . | nindent 8 }}
       {{- end }}
-      {{- with .Values.affinity }}
+      {{- with (.Values.api.affinity | default .Values.affinity) }}
       affinity:
         {{- toYaml . | nindent 8 }}
       {{- end }}

helm/hindsight/templates/controlplane-deployment.yaml

@@ -33,7 +33,7 @@ spec:
       - name: control-plane
         securityContext:
           {{- toYaml .Values.securityContext | nindent 10 }}
-        image: "{{ .Values.controlPlane.image.repository }}:{{ .Values.controlPlane.image.tag | default .Values.version }}"
+        image: "{{ .Values.controlPlane.image.repository }}:{{ .Values.controlPlane.image.tag | default .Values.version | default .Chart.AppVersion }}"
         imagePullPolicy: {{ .Values.controlPlane.image.pullPolicy }}
         ports:
         - name: http
@@ -71,7 +71,7 @@ spec:
       nodeSelector:
         {{- toYaml . | nindent 8 }}
       {{- end }}
-      {{- with .Values.affinity }}
+      {{- with (.Values.controlPlane.affinity | default .Values.affinity) }}
       affinity:
         {{- toYaml . | nindent 8 }}
       {{- end }}

helm/hindsight/templates/pdb.yaml

@@ -0,0 +1,56 @@
+{{- if and .Values.api.enabled .Values.api.podDisruptionBudget.enabled }}
+apiVersion: policy/v1
+kind: PodDisruptionBudget
+metadata:
+  name: {{ include "hindsight.fullname" . }}-api
+  labels:
+    {{- include "hindsight.api.labels" . | nindent 4 }}
+spec:
+  {{- if .Values.api.podDisruptionBudget.minAvailable }}
+  minAvailable: {{ .Values.api.podDisruptionBudget.minAvailable }}
+  {{- end }}
+  {{- if .Values.api.podDisruptionBudget.maxUnavailable }}
+  maxUnavailable: {{ .Values.api.podDisruptionBudget.maxUnavailable }}
+  {{- end }}
+  selector:
+    matchLabels:
+      {{- include "hindsight.api.selectorLabels" . | nindent 6 }}
+{{- end }}
+---
+{{- if and .Values.controlPlane.enabled .Values.controlPlane.podDisruptionBudget.enabled }}
+apiVersion: policy/v1
+kind: PodDisruptionBudget
+metadata:
+  name: {{ include "hindsight.fullname" . }}-control-plane
+  labels:
+    {{- include "hindsight.controlPlane.labels" . | nindent 4 }}
+spec:
+  {{- if .Values.controlPlane.podDisruptionBudget.minAvailable }}
+  minAvailable: {{ .Values.controlPlane.podDisruptionBudget.minAvailable }}
+  {{- end }}
+  {{- if .Values.controlPlane.podDisruptionBudget.maxUnavailable }}
+  maxUnavailable: {{ .Values.controlPlane.podDisruptionBudget.maxUnavailable }}
+  {{- end }}
+  selector:
+    matchLabels:
+      {{- include "hindsight.controlPlane.selectorLabels" . | nindent 6 }}
+{{- end }}
+---
+{{- if and .Values.worker.enabled .Values.worker.podDisruptionBudget.enabled }}
+apiVersion: policy/v1
+kind: PodDisruptionBudget
+metadata:
+  name: {{ include "hindsight.fullname" . }}-worker
+  labels:
+    {{- include "hindsight.worker.labels" . | nindent 4 }}
+spec:
+  {{- if .Values.worker.podDisruptionBudget.minAvailable }}
+  minAvailable: {{ .Values.worker.podDisruptionBudget.minAvailable }}
+  {{- end }}
+  {{- if .Values.worker.podDisruptionBudget.maxUnavailable }}
+  maxUnavailable: {{ .Values.worker.podDisruptionBudget.maxUnavailable }}
+  {{- end }}
+  selector:
+    matchLabels:
+      {{- include "hindsight.worker.selectorLabels" . | nindent 6 }}
+{{- end }}

helm/hindsight/templates/tei-embedding-deployment.yaml

@@ -0,0 +1,76 @@
+{{- if .Values.tei.embedding.enabled }}
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: {{ include "hindsight.fullname" . }}-tei-embedding
+  labels:
+    {{- include "hindsight.tei.embedding.labels" . | nindent 4 }}
+spec:
+  replicas: {{ .Values.tei.embedding.replicaCount }}
+  selector:
+    matchLabels:
+      {{- include "hindsight.tei.embedding.selectorLabels" . | nindent 6 }}
+  template:
+    metadata:
+      {{- with .Values.podAnnotations }}
+      annotations:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
+      labels:
+        {{- include "hindsight.tei.embedding.selectorLabels" . | nindent 8 }}
+    spec:
+      {{- if .Values.serviceAccount.create }}
+      serviceAccountName: {{ include "hindsight.serviceAccountName" . }}
+      {{- end }}
+      securityContext:
+        {{- toYaml .Values.podSecurityContext | nindent 8 }}
+      containers:
+      - name: tei-embedding
+        securityContext:
+          {{- toYaml .Values.securityContext | nindent 10 }}
+        image: "{{ .Values.tei.embedding.image.repository }}:{{ .Values.tei.embedding.image.tag }}"
+        imagePullPolicy: {{ .Values.tei.embedding.image.pullPolicy }}
+        args:
+          - "--model-id"
+          - {{ .Values.tei.embedding.model | quote }}
+          - "--hostname"
+          - "0.0.0.0"
+          {{- range .Values.tei.embedding.args }}
+          - {{ . | quote }}
+          {{- end }}
+        ports:
+        - name: http
+          containerPort: {{ .Values.tei.embedding.port }}
+          protocol: TCP
+        env:
+        - name: PORT
+          value: {{ .Values.tei.embedding.port | quote }}
+        {{- range $key, $value := .Values.tei.embedding.env }}
+        - name: {{ $key }}
+          value: {{ $value | quote }}
+        {{- end }}
+        livenessProbe:
+          {{- toYaml .Values.tei.embedding.livenessProbe | nindent 10 }}
+        readinessProbe:
+          {{- toYaml .Values.tei.embedding.readinessProbe | nindent 10 }}
+        resources:
+          {{- toYaml .Values.tei.embedding.resources | nindent 10 }}
+        volumeMounts:
+        - name: model-cache
+          mountPath: /data
+      volumes:
+      - name: model-cache
+        emptyDir: {}
+      {{- with .Values.nodeSelector }}
+      nodeSelector:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
+      {{- with .Values.affinity }}
+      affinity:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
+      {{- with .Values.tolerations }}
+      tolerations:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
+{{- end }}

helm/hindsight/templates/tei-embedding-service.yaml

@@ -0,0 +1,17 @@
+{{- if .Values.tei.embedding.enabled }}
+apiVersion: v1
+kind: Service
+metadata:
+  name: {{ include "hindsight.fullname" . }}-tei-embedding
+  labels:
+    {{- include "hindsight.tei.embedding.labels" . | nindent 4 }}
+spec:
+  type: ClusterIP
+  ports:
+  - port: {{ .Values.tei.embedding.port }}
+    targetPort: http
+    protocol: TCP
+    name: http
+  selector:
+    {{- include "hindsight.tei.embedding.selectorLabels" . | nindent 4 }}
+{{- end }}

helm/hindsight/templates/tei-reranker-deployment.yaml

@@ -0,0 +1,76 @@
+{{- if .Values.tei.reranker.enabled }}
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: {{ include "hindsight.fullname" . }}-tei-reranker
+  labels:
+    {{- include "hindsight.tei.reranker.labels" . | nindent 4 }}
+spec:
+  replicas: {{ .Values.tei.reranker.replicaCount }}
+  selector:
+    matchLabels:
+      {{- include "hindsight.tei.reranker.selectorLabels" . | nindent 6 }}
+  template:
+    metadata:
+      {{- with .Values.podAnnotations }}
+      annotations:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
+      labels:
+        {{- include "hindsight.tei.reranker.selectorLabels" . | nindent 8 }}
+    spec:
+      {{- if .Values.serviceAccount.create }}
+      serviceAccountName: {{ include "hindsight.serviceAccountName" . }}
+      {{- end }}
+      securityContext:
+        {{- toYaml .Values.podSecurityContext | nindent 8 }}
+      containers:
+      - name: tei-reranker
+        securityContext:
+          {{- toYaml .Values.securityContext | nindent 10 }}
+        image: "{{ .Values.tei.reranker.image.repository }}:{{ .Values.tei.reranker.image.tag }}"
+        imagePullPolicy: {{ .Values.tei.reranker.image.pullPolicy }}
+        args:
+          - "--model-id"
+          - {{ .Values.tei.reranker.model | quote }}
+          - "--hostname"
+          - "0.0.0.0"
+          {{- range .Values.tei.reranker.args }}
+          - {{ . | quote }}
+          {{- end }}
+        ports:
+        - name: http
+          containerPort: {{ .Values.tei.reranker.port }}
+          protocol: TCP
+        env:
+        - name: PORT
+          value: {{ .Values.tei.reranker.port | quote }}
+        {{- range $key, $value := .Values.tei.reranker.env }}
+        - name: {{ $key }}
+          value: {{ $value | quote }}
+        {{- end }}
+        livenessProbe:
+          {{- toYaml .Values.tei.reranker.livenessProbe | nindent 10 }}
+        readinessProbe:
+          {{- toYaml .Values.tei.reranker.readinessProbe | nindent 10 }}
+        resources:
+          {{- toYaml .Values.tei.reranker.resources | nindent 10 }}
+        volumeMounts:
+        - name: model-cache
+          mountPath: /data
+      volumes:
+      - name: model-cache
+        emptyDir: {}
+      {{- with .Values.nodeSelector }}
+      nodeSelector:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
+      {{- with .Values.affinity }}
+      affinity:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
+      {{- with .Values.tolerations }}
+      tolerations:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
+{{- end }}

helm/hindsight/templates/tei-reranker-service.yaml

@@ -0,0 +1,17 @@
+{{- if .Values.tei.reranker.enabled }}
+apiVersion: v1
+kind: Service
+metadata:
+  name: {{ include "hindsight.fullname" . }}-tei-reranker
+  labels:
+    {{- include "hindsight.tei.reranker.labels" . | nindent 4 }}
+spec:
+  type: ClusterIP
+  ports:
+  - port: {{ .Values.tei.reranker.port }}
+    targetPort: http
+    protocol: TCP
+    name: http
+  selector:
+    {{- include "hindsight.tei.reranker.selectorLabels" . | nindent 4 }}
+{{- end }}

helm/hindsight/templates/worker-service.yaml

@@ -0,0 +1,25 @@
+{{- if .Values.worker.enabled }}
+apiVersion: v1
+kind: Service
+metadata:
+  name: {{ include "hindsight.fullname" . }}-worker
+  labels:
+    {{- include "hindsight.worker.labels" . | nindent 4 }}
+  {{- if .Values.podAnnotations }}
+  annotations:
+    {{- /* Common Prometheus annotations for metrics scraping */}}
+    prometheus.io/scrape: "true"
+    prometheus.io/port: {{ .Values.worker.service.port | quote }}
+    prometheus.io/path: "/metrics"
+  {{- end }}
+spec:
+  # Headless service for StatefulSet (enables stable DNS names like worker-0.worker.namespace)
+  clusterIP: None
+  ports:
+  - port: {{ .Values.worker.service.port }}
+    targetPort: {{ .Values.worker.service.targetPort }}
+    protocol: TCP
+    name: http
+  selector:
+    {{- include "hindsight.worker.selectorLabels" . | nindent 4 }}
+{{- end }}

helm/hindsight/templates/worker-statefulset.yaml

@@ -0,0 +1,110 @@
+{{- if .Values.worker.enabled }}
+apiVersion: apps/v1
+kind: StatefulSet
+metadata:
+  name: {{ include "hindsight.fullname" . }}-worker
+  labels:
+    {{- include "hindsight.worker.labels" . | nindent 4 }}
+spec:
+  serviceName: {{ include "hindsight.fullname" . }}-worker
+  replicas: {{ .Values.worker.replicaCount }}
+  selector:
+    matchLabels:
+      {{- include "hindsight.worker.selectorLabels" . | nindent 6 }}
+  template:
+    metadata:
+      annotations:
+        {{- if not .Values.existingSecret }}
+        checksum/secret: {{ include (print $.Template.BasePath "/secret.yaml") . | sha256sum }}
+        {{- end }}
+        {{- with .Values.podAnnotations }}
+        {{- toYaml . | nindent 8 }}
+        {{- end }}
+      labels:
+        {{- include "hindsight.worker.selectorLabels" . | nindent 8 }}
+    spec:
+      {{- if .Values.serviceAccount.create }}
+      serviceAccountName: {{ include "hindsight.serviceAccountName" . }}
+      {{- end }}
+      securityContext:
+        {{- toYaml .Values.podSecurityContext | nindent 8 }}
+      containers:
+      - name: worker
+        securityContext:
+          {{- toYaml .Values.securityContext | nindent 10 }}
+        image: "{{ .Values.worker.image.repository }}:{{ .Values.worker.image.tag | default .Values.version | default .Chart.AppVersion }}"
+        imagePullPolicy: {{ .Values.worker.image.pullPolicy }}
+        command: ["hindsight-worker"]
+        ports:
+        - name: http
+          containerPort: {{ .Values.worker.service.targetPort }}
+          protocol: TCP
+        {{- if .Values.existingSecret }}
+        envFrom:
+        - secretRef:
+            name: {{ .Values.existingSecret }}
+        {{- end }}
+        env:
+        {{- /* POSTGRES_PASSWORD must be defined before DATABASE_URL for $(VAR) interpolation */}}
+        {{- if not .Values.postgresql.enabled }}
+        - name: POSTGRES_PASSWORD
+          valueFrom:
+            secretKeyRef:
+              name: {{ include "hindsight.secretName" . }}
+              key: postgres-password
+        {{- end }}
+        - name: HINDSIGHT_API_DATABASE_URL
+          value: {{ include "hindsight.databaseUrl" . | quote }}
+        {{- /* Worker ID uses pod name (StatefulSet provides stable names like worker-0, worker-1) */}}
+        - name: HINDSIGHT_API_WORKER_ID
+          valueFrom:
+            fieldRef:
+              fieldPath: metadata.name
+        {{- /* Inherit LLM config from api.env */}}
+        {{- range $key, $value := .Values.api.env }}
+        - name: {{ $key }}
+          value: {{ $value | quote }}
+        {{- end }}
+        {{- /* Worker-specific env vars */}}
+        {{- range $key, $value := .Values.worker.env }}
+        - name: {{ $key }}
+          value: {{ $value | quote }}
+        {{- end }}
+        {{- /* Only use secrets when not using existingSecret */}}
+        {{- if not .Values.existingSecret }}
+        {{- /* Inherit secrets from api.secrets */}}
+        {{- range $key, $value := .Values.api.secrets }}
+        - name: {{ $key }}
+          valueFrom:
+            secretKeyRef:
+              name: {{ include "hindsight.secretName" $ }}
+              key: {{ $key }}
+        {{- end }}
+        {{- /* Worker-specific secrets (can override api.secrets) */}}
+        {{- range $key, $value := .Values.worker.secrets }}
+        - name: {{ $key }}
+          valueFrom:
+            secretKeyRef:
+              name: {{ include "hindsight.secretName" $ }}
+              key: {{ $key }}
+        {{- end }}
+        {{- end }}
+        livenessProbe:
+          {{- toYaml .Values.worker.livenessProbe | nindent 10 }}
+        readinessProbe:
+          {{- toYaml .Values.worker.readinessProbe | nindent 10 }}
+        resources:
+          {{- toYaml .Values.worker.resources | nindent 10 }}
+      {{- with .Values.nodeSelector }}
+      nodeSelector:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
+      {{- with (.Values.worker.affinity | default .Values.affinity) }}
+      affinity:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
+      {{- with .Values.tolerations }}
+      tolerations:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
+{{- end }}

helm/hindsight/values.yaml

@@ -1,7 +1,8 @@
 # Default values for hindsight
 
-# Chart version - use this to set a consistent image tag across all components
-version: "0.1.1"
+# Global version override - use this to set a consistent image tag across all components
+# If not set, defaults to Chart.appVersion from Chart.yaml
+# version: ""
 
 # Use an existing secret instead of creating one from values
 # When set, all keys from this secret are injected as environment variables via envFrom
@@ -57,6 +58,15 @@ api:
     timeoutSeconds: 3
     failureThreshold: 3
 
+  # Pod disruption budget
+  podDisruptionBudget:
+    enabled: false
+    minAvailable: 1
+    # maxUnavailable: 1
+
+  # Pod affinity/anti-affinity (overrides global affinity for this component)
+  # affinity: {}
+
   # Environment variables
   env:
     #HINDSIGHT_API_LLM_PROVIDER: "groq"
@@ -67,6 +77,72 @@ api:
     # HINDSIGHT_API_LLM_API_KEY: "your-api-key"
     # HINDSIGHT_API_LLM_BASE_URL: "https://api.groq.com/openai/v1"
 
+# Worker settings (distributed task processing)
+# When enabled, dedicated worker pods process tasks and the API's internal worker is disabled
+worker:
+  enabled: false
+  replicaCount: 2
+  image:
+    repository: ghcr.io/vectorize-io/hindsight-api
+    pullPolicy: IfNotPresent
+    # tag: ""  # defaults to .Values.version, then Chart.appVersion if not specified
+
+  service:
+    # Service for metrics scraping (headless for StatefulSet)
+    port: 8889
+    targetPort: 8889
+
+  # Resource limits and requests
+  resources:
+    limits:
+      cpu: 2000m
+      memory: 4Gi
+    requests:
+      cpu: 500m
+      memory: 1Gi
+
+  # Liveness and readiness probes
+  livenessProbe:
+    httpGet:
+      path: /health
+      port: 8889
+    initialDelaySeconds: 30
+    periodSeconds: 10
+    timeoutSeconds: 5
+    failureThreshold: 3
+
+  readinessProbe:
+    httpGet:
+      path: /health
+      port: 8889
+    initialDelaySeconds: 10
+    periodSeconds: 5
+    timeoutSeconds: 3
+    failureThreshold: 3
+
+  # Worker-specific environment variables
+  env:
+    # Poll interval in milliseconds (how often to check for new tasks)
+    HINDSIGHT_API_WORKER_POLL_INTERVAL_MS: "500"
+    # Number of tasks to claim per poll cycle
+    HINDSIGHT_API_WORKER_BATCH_SIZE: "10"
+    # Max retries before marking a task as failed
+    HINDSIGHT_API_WORKER_MAX_RETRIES: "3"
+    # HTTP port for metrics/health (matches service.targetPort)
+    HINDSIGHT_API_WORKER_HTTP_PORT: "8889"
+
+  # Pod disruption budget
+  podDisruptionBudget:
+    enabled: false
+    minAvailable: 1
+    # maxUnavailable: 1
+
+  # Pod affinity/anti-affinity (overrides global affinity for this component)
+  # affinity: {}
+
+  # Secret environment variables (inherited from api.secrets if not specified)
+  secrets: {}
+
 # Image settings for control plane
 controlPlane:
   enabled: true
@@ -107,6 +183,15 @@ controlPlane:
     timeoutSeconds: 3
     failureThreshold: 3
 
+  # Pod disruption budget
+  podDisruptionBudget:
+    enabled: false
+    minAvailable: 1
+    # maxUnavailable: 1
+
+  # Pod affinity/anti-affinity (overrides global affinity for this component)
+  # affinity: {}
+
   # Environment variables
   env:
     NODE_ENV: "production"
@@ -205,9 +290,87 @@ nodeSelector: {}
 # Tolerations
 tolerations: []
 
-# Affinity
+# Affinity (applied to all components unless overridden per-component)
 affinity: {}
 
+# TEI (Text Embeddings Inference) - optional standalone deployments
+# for reranking and/or embedding models
+tei:
+  reranker:
+    enabled: false
+    replicaCount: 1
+    image:
+      repository: ghcr.io/huggingface/text-embeddings-inference
+      tag: cpu-1.8.3
+      pullPolicy: IfNotPresent
+    model: "cross-encoder/ms-marco-MiniLM-L-6-v2"
+    port: 8090
+    args:
+      - "--auto-truncate"
+    env:
+      PAYLOAD_LIMIT: "10000000"
+      MAX_CLIENT_BATCH_SIZE: "256"
+    resources:
+      limits:
+        cpu: 2000m
+        memory: 2Gi
+      requests:
+        cpu: 500m
+        memory: 1Gi
+    livenessProbe:
+      httpGet:
+        path: /health
+        port: 8090
+      initialDelaySeconds: 30
+      periodSeconds: 10
+      timeoutSeconds: 5
+      failureThreshold: 6
+    readinessProbe:
+      httpGet:
+        path: /health
+        port: 8090
+      initialDelaySeconds: 15
+      periodSeconds: 5
+      timeoutSeconds: 3
+      failureThreshold: 3
+
+  embedding:
+    enabled: false
+    replicaCount: 1
+    image:
+      repository: ghcr.io/huggingface/text-embeddings-inference
+      tag: cpu-1.8.3
+      pullPolicy: IfNotPresent
+    model: "sentence-transformers/all-MiniLM-L6-v2"
+    port: 8091
+    args: []
+    env:
+      PAYLOAD_LIMIT: "10000000"
+      MAX_CLIENT_BATCH_SIZE: "256"
+    resources:
+      limits:
+        cpu: 2000m
+        memory: 2Gi
+      requests:
+        cpu: 500m
+        memory: 1Gi
+    livenessProbe:
+      httpGet:
+        path: /health
+        port: 8091
+      initialDelaySeconds: 30
+      periodSeconds: 10
+      timeoutSeconds: 5
+      failureThreshold: 6
+    readinessProbe:
+      httpGet:
+        path: /health
+        port: 8091
+      initialDelaySeconds: 15
+      periodSeconds: 5
+      timeoutSeconds: 3
+      failureThreshold: 3
+
 # Autoscaling
 autoscaling:
   enabled: false

hindsight-api/hindsight_api/__init__.py

@@ -46,4 +46,4 @@ __all__ = [
     "RemoteTEICrossEncoder",
     "LLMConfig",
 ]
-__version__ = "0.1.0"
+__version__ = "0.4.11"

hindsight-api/hindsight_api/admin/cli.py

@@ -244,6 +244,65 @@ def run_db_migration(
     typer.echo("Database migrations completed successfully")
 
 
+async def _decommission_worker(db_url: str, worker_id: str, schema: str = "public") -> int:
+    """Release all tasks owned by a worker, setting them back to pending status."""
+    is_pg0, instance_name, _ = parse_pg0_url(db_url)
+    if is_pg0:
+        typer.echo(f"Starting embedded PostgreSQL (instance: {instance_name})...")
+    resolved_url = await resolve_database_url(db_url)
+
+    conn = await asyncpg.connect(resolved_url)
+    try:
+        table = _fq_table("async_operations", schema)
+        result = await conn.fetch(
+            f"""
+            UPDATE {table}
+            SET status = 'pending', worker_id = NULL, claimed_at = NULL, updated_at = now()
+            WHERE worker_id = $1 AND status = 'processing'
+            RETURNING operation_id
+            """,
+            worker_id,
+        )
+        return len(result)
+    finally:
+        await conn.close()
+
+
+@app.command(name="decommission-worker")
+def decommission_worker(
+    worker_id: str = typer.Argument(..., help="Worker ID to decommission"),
+    schema: str = typer.Option("public", "--schema", "-s", help="Database schema"),
+    yes: bool = typer.Option(False, "--yes", "-y", help="Skip confirmation prompt"),
+):
+    """Release all tasks owned by a worker (sets status back to pending).
+
+    Use this command when a worker has crashed or been removed without graceful shutdown.
+    All tasks that were being processed by the worker will be released back to the queue
+    so other workers can pick them up.
+    """
+    config = HindsightConfig.from_env()
+
+    if not config.database_url:
+        typer.echo("Error: Database URL not configured.", err=True)
+        typer.echo("Set HINDSIGHT_API_DATABASE_URL environment variable.", err=True)
+        raise typer.Exit(1)
+
+    if not yes:
+        typer.confirm(
+            f"This will release all tasks owned by worker '{worker_id}' back to pending. Continue?",
+            abort=True,
+        )
+
+    typer.echo(f"Decommissioning worker '{worker_id}' (schema: {schema})...")
+
+    count = asyncio.run(_decommission_worker(config.database_url, worker_id, schema))
+
+    if count > 0:
+        typer.echo(f"Released {count} task(s) from worker '{worker_id}'")
+    else:
+        typer.echo(f"No tasks found for worker '{worker_id}'")
+
+
 def main():
     app()
 

hindsight-api/hindsight_api/alembic/versions/5a366d414dce_initial_schema.py

@@ -6,11 +6,13 @@ Create Date: 2025-11-27 11:54:19.228030
 
 """
 
+import os
 from collections.abc import Sequence
 
 import sqlalchemy as sa
 from alembic import op
 from pgvector.sqlalchemy import Vector
+from sqlalchemy import text
 from sqlalchemy.dialects import postgresql
 
 # revision identifiers, used by Alembic.
@@ -20,11 +22,114 @@ branch_labels: str | Sequence[str] | None = None
 depends_on: str | Sequence[str] | None = None
 
 
+def _detect_vector_extension() -> str:
+    """
+    Detect or validate vector extension: 'pgvector', 'vchord', or 'pgvectorscale'.
+    Respects HINDSIGHT_API_VECTOR_EXTENSION env var if set.
+    """
+    conn = op.get_bind()
+    vector_extension = os.getenv("HINDSIGHT_API_VECTOR_EXTENSION", "pgvector").lower()
+
+    # Validate configured extension is installed
+    if vector_extension == "pgvectorscale":
+        # pgvectorscale/DiskANN requires pgvector
+        pgvector_check = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vector'")).scalar()
+        if not pgvector_check:
+            raise RuntimeError(
+                "DiskANN requires pgvector. Install with: CREATE EXTENSION vector; then vectorscale or pg_diskann CASCADE;"
+            )
+        # Check for either vectorscale (open source) or pg_diskann (Azure)
+        vectorscale_check = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vectorscale'")).scalar()
+        pg_diskann_check = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'pg_diskann'")).scalar()
+
+        if vectorscale_check:
+            return "pgvectorscale"
+        elif pg_diskann_check:
+            return "pg_diskann"
+        else:
+            raise RuntimeError(
+                "Configured vector extension 'pgvectorscale' not found. Install either:\n"
+                "  - pgvectorscale: CREATE EXTENSION vectorscale CASCADE;\n"
+                "  - pg_diskann (Azure): CREATE EXTENSION pg_diskann CASCADE;"
+            )
+    elif vector_extension == "vchord":
+        vchord_check = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vchord'")).scalar()
+        if not vchord_check:
+            raise RuntimeError(
+                "Configured vector extension 'vchord' not found. Install it with: CREATE EXTENSION vchord CASCADE;"
+            )
+        return "vchord"
+    elif vector_extension == "pgvector":
+        pgvector_check = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vector'")).scalar()
+        if not pgvector_check:
+            raise RuntimeError(
+                "Configured vector extension 'pgvector' not found. Install it with: CREATE EXTENSION vector;"
+            )
+        return "pgvector"
+    else:
+        raise ValueError(
+            f"Invalid HINDSIGHT_API_VECTOR_EXTENSION: {vector_extension}. Must be 'pgvector', 'vchord', or 'pgvectorscale'"
+        )
+
+
+def _detect_text_search_extension() -> str:
+    """
+    Detect or validate text search extension: 'native', 'vchord', or 'pg_textsearch'.
+    Respects HINDSIGHT_API_TEXT_SEARCH_EXTENSION env var.
+    Creates the extension if needed.
+    """
+    text_search_extension = os.getenv("HINDSIGHT_API_TEXT_SEARCH_EXTENSION", "native").lower()
+
+    if text_search_extension == "vchord":
+        # Create vchord_bm25 extension if not exists
+        try:
+            op.execute("CREATE EXTENSION IF NOT EXISTS vchord_bm25 CASCADE")
+        except Exception:
+            # Extension might already exist or user lacks permissions - verify it exists
+            conn = op.get_bind()
+            result = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vchord_bm25'")).fetchone()
+            if not result:
+                # Extension truly doesn't exist - re-raise the error
+                raise
+        return "vchord"
+    elif text_search_extension == "pg_textsearch":
+        # Create pg_textsearch extension if not exists
+        try:
+            op.execute("CREATE EXTENSION IF NOT EXISTS pg_textsearch CASCADE")
+        except Exception:
+            # Extension might already exist or user lacks permissions - verify it exists
+            conn = op.get_bind()
+            result = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'pg_textsearch'")).fetchone()
+            if not result:
+                # Extension truly doesn't exist - re-raise the error
+                raise
+        return "pg_textsearch"
+    elif text_search_extension == "native":
+        return "native"
+    else:
+        raise ValueError(
+            f"Invalid HINDSIGHT_API_TEXT_SEARCH_EXTENSION: {text_search_extension}. Must be 'native', 'vchord', or 'pg_textsearch'"
+        )
+
+
 def upgrade() -> None:
     """Upgrade schema - create all tables from scratch."""
 
-    # Enable required extensions
-    op.execute("CREATE EXTENSION IF NOT EXISTS vector")
+    # Note: pgvector extension is installed globally BEFORE migrations run
+    # See migrations.py:run_migrations() - this ensures the extension is available
+    # to all schemas, not just the one being migrated
+
+    # We keep this here as a fallback for backwards compatibility
+    # This may fail if user lacks permissions, which is fine if extension already exists
+    try:
+        op.execute("CREATE EXTENSION IF NOT EXISTS vector")
+    except Exception:
+        # Extension might already exist or user lacks permissions - verify it exists
+        conn = op.get_bind()
+        result = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vector'")).fetchone()
+        if not result:
+            # Extension truly doesn't exist - re-raise the error
+            raise
 
     # Create banks table
     op.create_table(
@@ -152,11 +257,29 @@ def upgrade() -> None:
     )
 
     # Add search_vector column for full-text search
-    op.execute("""
-        ALTER TABLE memory_units
-        ADD COLUMN search_vector tsvector
-        GENERATED ALWAYS AS (to_tsvector('english', COALESCE(text, '') || ' ' || COALESCE(context, ''))) STORED
-    """)
+    # Type depends on configured text search backend
+    text_search_ext = _detect_text_search_extension()
+
+    if text_search_ext == "vchord":
+        # VectorChord BM25: bm25vector type (no GENERATED - tokenization happens on INSERT)
+        # Note: vchord_bm25 extension creates types in bm25_catalog schema
+        op.execute("""
+            ALTER TABLE memory_units
+            ADD COLUMN search_vector bm25_catalog.bm25vector
+        """)
+    elif text_search_ext == "pg_textsearch":
+        # Timescale pg_textsearch: dummy TEXT column for consistency (indexes operate on base columns directly)
+        op.execute("""
+            ALTER TABLE memory_units
+            ADD COLUMN search_vector TEXT
+        """)
+    else:  # native
+        # Native PostgreSQL: tsvector with automatic generation
+        op.execute("""
+            ALTER TABLE memory_units
+            ADD COLUMN search_vector tsvector
+            GENERATED ALWAYS AS (to_tsvector('english', COALESCE(text, '') || ' ' || COALESCE(context, ''))) STORED
+        """)
 
     op.create_index("idx_memory_units_bank_id", "memory_units", ["bank_id"])
     op.create_index("idx_memory_units_document_id", "memory_units", ["document_id"])
@@ -186,19 +309,61 @@ def upgrade() -> None:
         ["bank_id", sa.text("event_date DESC")],
         postgresql_where=sa.text("fact_type = 'observation'"),
     )
-    op.create_index(
-        "idx_memory_units_embedding",
-        "memory_units",
-        ["embedding"],
-        postgresql_using="hnsw",
-        postgresql_ops={"embedding": "vector_cosine_ops"},
-    )
+    # Create vector index - conditional based on available extension
+    vector_ext = _detect_vector_extension()
 
-    # Create BM25 full-text search index on search_vector
-    op.execute("""
-        CREATE INDEX idx_memory_units_text_search ON memory_units
-        USING gin(search_vector)
-    """)
+    if vector_ext == "pgvectorscale":
+        # Use DiskANN index for pgvectorscale (disk-based, scalable)
+        op.execute("""
+            CREATE INDEX idx_memory_units_embedding ON memory_units
+            USING diskann (embedding vector_cosine_ops)
+            WITH (num_neighbors = 50)
+        """)
+    elif vector_ext == "pg_diskann":
+        # Use DiskANN index for pg_diskann (Azure)
+        op.execute("""
+            CREATE INDEX idx_memory_units_embedding ON memory_units
+            USING diskann (embedding vector_cosine_ops)
+            WITH (max_neighbors = 50)
+        """)
+    elif vector_ext == "vchord":
+        # Use vchordrq index for vchord (supports high-dimensional embeddings)
+        op.execute("""
+            CREATE INDEX idx_memory_units_embedding ON memory_units
+            USING vchordrq (embedding vector_l2_ops)
+        """)
+    else:  # pgvector
+        # Use HNSW index for pgvector
+        op.create_index(
+            "idx_memory_units_embedding",
+            "memory_units",
+            ["embedding"],
+            postgresql_using="hnsw",
+            postgresql_ops={"embedding": "vector_cosine_ops"},
+        )
+
+    # Create full-text search index on search_vector
+    # Index type depends on text search backend
+    if text_search_ext == "vchord":
+        # VectorChord BM25 index
+        op.execute("""
+            CREATE INDEX idx_memory_units_text_search ON memory_units
+            USING bm25 (search_vector bm25_catalog.bm25_ops)
+        """)
+    elif text_search_ext == "pg_textsearch":
+        # Timescale pg_textsearch BM25 index on text column
+        # Note: pg_textsearch doesn't support expressions, so we index the main text column
+        op.execute("""
+            CREATE INDEX idx_memory_units_text_search ON memory_units
+            USING bm25(text)
+            WITH (text_config='english')
+        """)
+    else:  # native
+        # Native PostgreSQL GIN index
+        op.execute("""
+            CREATE INDEX idx_memory_units_text_search ON memory_units
+            USING gin(search_vector)
+        """)
 
     op.execute("""
         CREATE MATERIALIZED VIEW memory_units_bm25 AS

hindsight-api/hindsight_api/alembic/versions/a1b2c3d4e5f6_add_file_storage_table.py

@@ -0,0 +1,70 @@
+"""Add file_storage table for BYTEA-based file storage
+
+Revision ID: a1b2c3d4e5f6
+Revises: y0t1u2v3w4x5
+Create Date: 2026-02-16
+
+Creates a dedicated table for storing uploaded files using BYTEA.
+This provides zero-config file storage that "just works" for development
+and small deployments. For production/scale, use S3-compatible storage.
+
+Files are stored in a separate table to avoid bloating the documents table.
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+revision: str = "a1b2c3d4e5f6"
+down_revision: str | Sequence[str] | None = "y0t1u2v3w4x5"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (required for multi-tenant support)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    """Create file_storage table for BYTEA storage."""
+    schema = _get_schema_prefix()
+
+    # Create file_storage table (minimal: just key + data)
+    op.execute(
+        f"""
+        CREATE TABLE {schema}file_storage (
+            storage_key TEXT PRIMARY KEY,
+            data BYTEA NOT NULL
+        )
+    """
+    )
+
+    # Add file tracking columns to documents table
+    op.execute(
+        f"""
+        ALTER TABLE {schema}documents
+        ADD COLUMN IF NOT EXISTS file_storage_key TEXT,
+        ADD COLUMN IF NOT EXISTS file_original_name TEXT,
+        ADD COLUMN IF NOT EXISTS file_content_type TEXT
+    """
+    )
+
+
+def downgrade() -> None:
+    """Remove file_storage table and related columns."""
+    schema = _get_schema_prefix()
+
+    # Drop columns from documents table
+    op.execute(
+        f"""
+        ALTER TABLE {schema}documents
+        DROP COLUMN IF EXISTS file_storage_key,
+        DROP COLUMN IF EXISTS file_original_name,
+        DROP COLUMN IF EXISTS file_content_type
+    """
+    )
+
+    # Drop file_storage table
+    op.execute(f"DROP TABLE IF EXISTS {schema}file_storage")

hindsight-api/hindsight_api/alembic/versions/f1a2b3c4d5e6_add_memory_links_composite_index.py

@@ -0,0 +1,44 @@
+"""add_memory_links_from_type_weight_index
+
+Revision ID: f1a2b3c4d5e6
+Revises: e0a1b2c3d4e5
+Create Date: 2025-01-12
+
+Add composite index on memory_links (from_unit_id, link_type, weight DESC)
+to optimize MPFP graph traversal queries that need top-k edges per type.
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+# revision identifiers, used by Alembic.
+revision: str = "f1a2b3c4d5e6"
+down_revision: str | Sequence[str] | None = "e0a1b2c3d4e5"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (e.g., 'tenant_x.' or '' for public)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    """Add composite index for efficient MPFP edge loading."""
+    schema = _get_schema_prefix()
+    # Create composite index for efficient top-k per (from_node, link_type) queries
+    # This enables LATERAL joins to use index-only scans with early termination
+    # Note: Not using CONCURRENTLY here as it requires running outside a transaction
+    # For production with large tables, consider running this manually with CONCURRENTLY
+    op.execute(
+        f"CREATE INDEX IF NOT EXISTS idx_memory_links_from_type_weight "
+        f"ON {schema}memory_links(from_unit_id, link_type, weight DESC)"
+    )
+
+
+def downgrade() -> None:
+    """Remove the composite index."""
+    schema = _get_schema_prefix()
+    op.execute(f"DROP INDEX IF EXISTS {schema}idx_memory_links_from_type_weight")

hindsight-api/hindsight_api/alembic/versions/g2a3b4c5d6e7_add_tags_column.py

@@ -0,0 +1,48 @@
+"""add_tags_column
+
+Revision ID: g2a3b4c5d6e7
+Revises: f1a2b3c4d5e6
+Create Date: 2025-01-13
+
+Add tags column to memory_units and documents tables for visibility scoping.
+Tags enable filtering memories by scope (e.g., user IDs, session IDs) during recall/reflect.
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+# revision identifiers, used by Alembic.
+revision: str = "g2a3b4c5d6e7"
+down_revision: str | Sequence[str] | None = "f1a2b3c4d5e6"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (e.g., 'tenant_x.' or '' for public)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    """Add tags column to memory_units and documents tables."""
+    schema = _get_schema_prefix()
+
+    # Add tags column to memory_units table
+    op.execute(f"ALTER TABLE {schema}memory_units ADD COLUMN IF NOT EXISTS tags VARCHAR[] NOT NULL DEFAULT '{{}}'")
+
+    # Create GIN index for efficient array containment queries (tags && ARRAY['x'])
+    op.execute(f"CREATE INDEX IF NOT EXISTS idx_memory_units_tags ON {schema}memory_units USING GIN (tags)")
+
+    # Add tags column to documents table for document-level tags
+    op.execute(f"ALTER TABLE {schema}documents ADD COLUMN IF NOT EXISTS tags VARCHAR[] NOT NULL DEFAULT '{{}}'")
+
+
+def downgrade() -> None:
+    """Remove tags columns and index."""
+    schema = _get_schema_prefix()
+
+    op.execute(f"DROP INDEX IF EXISTS {schema}idx_memory_units_tags")
+    op.execute(f"ALTER TABLE {schema}memory_units DROP COLUMN IF EXISTS tags")
+    op.execute(f"ALTER TABLE {schema}documents DROP COLUMN IF EXISTS tags")

hindsight-api/hindsight_api/alembic/versions/h3c4d5e6f7g8_mental_models_v4.py

@@ -0,0 +1,112 @@
+"""mental_models_v4
+
+Revision ID: h3c4d5e6f7g8
+Revises: g2a3b4c5d6e7
+Create Date: 2026-01-08 00:00:00.000000
+
+This migration implements the v4 mental models system:
+1. Deletes existing observation memory_units (observations now in mental models)
+2. Adds mission column to banks (replacing background)
+3. Creates mental_models table with final schema
+
+Mental models can reference entities when an entity is "promoted" to a mental model.
+Summary content is stored as JSONB observations with per-observation fact attribution.
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+# revision identifiers, used by Alembic.
+revision: str = "h3c4d5e6f7g8"
+down_revision: str | Sequence[str] | None = "g2a3b4c5d6e7"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (required for multi-tenant support)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    """Apply mental models v4 changes."""
+    schema = _get_schema_prefix()
+
+    # Step 1: Delete observation memory_units (cascades to unit_entities links)
+    # Observations are now handled through mental models, not memory_units
+    op.execute(f"DELETE FROM {schema}memory_units WHERE fact_type = 'observation'")
+
+    # Step 2: Drop observation-specific index (if it exists)
+    op.execute(f"DROP INDEX IF EXISTS {schema}idx_memory_units_observation_date")
+
+    # Step 3: Add mission column to banks (replacing background)
+    op.execute(f"ALTER TABLE {schema}banks ADD COLUMN IF NOT EXISTS mission TEXT")
+
+    # Migrate: copy background to mission if background column exists
+    # Use DO block to check column existence first (idempotent for re-runs)
+    schema_name = context.config.get_main_option("target_schema") or "public"
+    op.execute(f"""
+        DO $$
+        BEGIN
+            IF EXISTS (
+                SELECT 1 FROM information_schema.columns
+                WHERE table_schema = '{schema_name}' AND table_name = 'banks' AND column_name = 'background'
+            ) THEN
+                UPDATE {schema}banks
+                SET mission = background
+                WHERE mission IS NULL;
+            END IF;
+        END $$;
+    """)
+
+    # Remove background column (replaced by mission)
+    op.execute(f"ALTER TABLE {schema}banks DROP COLUMN IF EXISTS background")
+
+    # Step 4: Create mental_models table with final v4 schema (if not exists)
+    op.execute(f"""
+        CREATE TABLE IF NOT EXISTS {schema}mental_models (
+            id VARCHAR(64) NOT NULL,
+            bank_id VARCHAR(64) NOT NULL,
+            subtype VARCHAR(32) NOT NULL,
+            name VARCHAR(256) NOT NULL,
+            description TEXT NOT NULL,
+            entity_id UUID,
+            observations JSONB DEFAULT '{{"observations": []}}'::jsonb,
+            links VARCHAR[],
+            tags VARCHAR[] DEFAULT '{{}}',
+            last_updated TIMESTAMP WITH TIME ZONE,
+            created_at TIMESTAMP WITH TIME ZONE NOT NULL DEFAULT now(),
+            PRIMARY KEY (id, bank_id),
+            FOREIGN KEY (bank_id) REFERENCES {schema}banks(bank_id) ON DELETE CASCADE,
+            FOREIGN KEY (entity_id) REFERENCES {schema}entities(id) ON DELETE SET NULL,
+            CONSTRAINT ck_mental_models_subtype CHECK (subtype IN ('structural', 'emergent', 'pinned', 'learned'))
+        )
+    """)
+
+    # Step 5: Create indexes for efficient queries (if not exist)
+    op.execute(f"CREATE INDEX IF NOT EXISTS idx_mental_models_bank_id ON {schema}mental_models(bank_id)")
+    op.execute(f"CREATE INDEX IF NOT EXISTS idx_mental_models_subtype ON {schema}mental_models(bank_id, subtype)")
+    op.execute(f"CREATE INDEX IF NOT EXISTS idx_mental_models_entity_id ON {schema}mental_models(entity_id)")
+    # GIN index for efficient tags array filtering
+    op.execute(f"CREATE INDEX IF NOT EXISTS idx_mental_models_tags ON {schema}mental_models USING GIN(tags)")
+
+
+def downgrade() -> None:
+    """Revert mental models v4 changes."""
+    schema = _get_schema_prefix()
+
+    # Drop mental_models table (cascades to indexes)
+    op.execute(f"DROP TABLE IF EXISTS {schema}mental_models CASCADE")
+
+    # Add back background column to banks
+    op.execute(f"ALTER TABLE {schema}banks ADD COLUMN IF NOT EXISTS background TEXT")
+
+    # Migrate mission back to background
+    op.execute(f"UPDATE {schema}banks SET background = mission WHERE background IS NULL")
+
+    # Remove mission column
+    op.execute(f"ALTER TABLE {schema}banks DROP COLUMN IF EXISTS mission")
+
+    # Note: Cannot restore deleted observations - they are lost on downgrade

hindsight-api/hindsight_api/alembic/versions/i4d5e6f7g8h9_delete_opinions.py

@@ -0,0 +1,41 @@
+"""delete_opinions
+
+Revision ID: i4d5e6f7g8h9
+Revises: h3c4d5e6f7g8
+Create Date: 2026-01-15 00:00:00.000000
+
+This migration removes opinion facts from memory_units.
+Opinions are no longer a separate fact type - they are now represented
+through mental model observations with confidence scores.
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+# revision identifiers, used by Alembic.
+revision: str = "i4d5e6f7g8h9"
+down_revision: str | Sequence[str] | None = "h3c4d5e6f7g8"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (required for multi-tenant support)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    """Delete opinion memory_units."""
+    schema = _get_schema_prefix()
+
+    # Delete opinion memory_units (cascades to unit_entities links)
+    # Opinions are now handled through mental model observations
+    op.execute(f"DELETE FROM {schema}memory_units WHERE fact_type = 'opinion'")
+
+
+def downgrade() -> None:
+    """Cannot restore deleted opinions."""
+    # Note: Cannot restore deleted opinions - they are lost on downgrade
+    pass

hindsight-api/hindsight_api/alembic/versions/j5e6f7g8h9i0_mental_model_versions.py

@@ -0,0 +1,95 @@
+"""mental_model_versions
+
+Revision ID: j5e6f7g8h9i0
+Revises: i4d5e6f7g8h9
+Create Date: 2026-01-16 00:00:00.000000
+
+This migration adds versioning support for mental models:
+1. Creates mental_model_versions table to store observation snapshots
+2. Adds version column to mental_models for tracking current version
+
+This enables changelog/diff functionality for mental model observations.
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+# revision identifiers, used by Alembic.
+revision: str = "j5e6f7g8h9i0"
+down_revision: str | Sequence[str] | None = "i4d5e6f7g8h9"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (required for multi-tenant support)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    """Create mental_model_versions table and add version tracking."""
+    schema = _get_schema_prefix()
+
+    # Create mental_model_versions table for storing observation snapshots
+    op.execute(f"""
+        CREATE TABLE {schema}mental_model_versions (
+            id SERIAL PRIMARY KEY,
+            mental_model_id VARCHAR(64) NOT NULL,
+            bank_id VARCHAR(64) NOT NULL,
+            version INT NOT NULL,
+            observations JSONB NOT NULL DEFAULT '{{"observations": []}}'::jsonb,
+            created_at TIMESTAMP WITH TIME ZONE NOT NULL DEFAULT now(),
+            FOREIGN KEY (mental_model_id, bank_id)
+                REFERENCES {schema}mental_models(id, bank_id) ON DELETE CASCADE,
+            UNIQUE (mental_model_id, bank_id, version)
+        )
+    """)
+
+    # Index for efficient version queries (get latest, list versions)
+    op.execute(f"""
+        CREATE INDEX idx_mental_model_versions_lookup
+        ON {schema}mental_model_versions(mental_model_id, bank_id, version DESC)
+    """)
+
+    # Add version column to mental_models to track current version
+    op.execute(f"""
+        ALTER TABLE {schema}mental_models
+        ADD COLUMN IF NOT EXISTS version INT NOT NULL DEFAULT 0
+    """)
+
+    # Migrate existing mental models: create version 1 for any that have observations
+    op.execute(f"""
+        INSERT INTO {schema}mental_model_versions (mental_model_id, bank_id, version, observations, created_at)
+        SELECT id, bank_id, 1, observations, COALESCE(last_updated, created_at)
+        FROM {schema}mental_models
+        WHERE observations IS NOT NULL
+          AND observations != '{{"observations": []}}'::jsonb
+          AND (observations->'observations') IS NOT NULL
+          AND jsonb_array_length(observations->'observations') > 0
+    """)
+
+    # Update version to 1 for migrated mental models
+    op.execute(f"""
+        UPDATE {schema}mental_models
+        SET version = 1
+        WHERE observations IS NOT NULL
+          AND observations != '{{"observations": []}}'::jsonb
+          AND (observations->'observations') IS NOT NULL
+          AND jsonb_array_length(observations->'observations') > 0
+    """)
+
+
+def downgrade() -> None:
+    """Remove mental_model_versions table and version column."""
+    schema = _get_schema_prefix()
+
+    # Drop index
+    op.execute(f"DROP INDEX IF EXISTS {schema}idx_mental_model_versions_lookup")
+
+    # Drop versions table
+    op.execute(f"DROP TABLE IF EXISTS {schema}mental_model_versions")
+
+    # Remove version column from mental_models
+    op.execute(f"ALTER TABLE {schema}mental_models DROP COLUMN IF EXISTS version")

hindsight-api/hindsight_api/alembic/versions/k6f7g8h9i0j1_add_directive_subtype.py

@@ -0,0 +1,58 @@
+"""add_directive_subtype
+
+Revision ID: k6f7g8h9i0j1
+Revises: j5e6f7g8h9i0
+Create Date: 2026-01-16 00:00:00.000000
+
+This migration adds 'directive' to the mental_models subtype constraint.
+Directives are hard rules with user-provided observations that the reflect agent must follow.
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+# revision identifiers, used by Alembic.
+revision: str = "k6f7g8h9i0j1"
+down_revision: str | Sequence[str] | None = "j5e6f7g8h9i0"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (required for multi-tenant support)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    """Add 'directive' to mental_models subtype constraint."""
+    schema = _get_schema_prefix()
+
+    # Drop existing constraint
+    op.execute(f"ALTER TABLE {schema}mental_models DROP CONSTRAINT IF EXISTS ck_mental_models_subtype")
+
+    # Create new constraint with 'directive' added
+    op.execute(f"""
+        ALTER TABLE {schema}mental_models
+        ADD CONSTRAINT ck_mental_models_subtype
+        CHECK (subtype IN ('structural', 'emergent', 'pinned', 'learned', 'directive'))
+    """)
+
+
+def downgrade() -> None:
+    """Remove 'directive' from mental_models subtype constraint."""
+    schema = _get_schema_prefix()
+
+    # First delete any directives (cannot downgrade if they exist)
+    op.execute(f"DELETE FROM {schema}mental_models WHERE subtype = 'directive'")
+
+    # Drop constraint with directive
+    op.execute(f"ALTER TABLE {schema}mental_models DROP CONSTRAINT IF EXISTS ck_mental_models_subtype")
+
+    # Recreate original constraint without directive
+    op.execute(f"""
+        ALTER TABLE {schema}mental_models
+        ADD CONSTRAINT ck_mental_models_subtype
+        CHECK (subtype IN ('structural', 'emergent', 'pinned', 'learned'))
+    """)

hindsight-api/hindsight_api/alembic/versions/l7g8h9i0j1k2_add_worker_columns.py

@@ -0,0 +1,109 @@
+"""add_worker_columns
+
+Revision ID: l7g8h9i0j1k2
+Revises: k6f7g8h9i0j1
+Create Date: 2026-01-19 00:00:00.000000
+
+This migration adds columns to async_operations for distributed worker support:
+- worker_id: ID of the worker that claimed the task
+- claimed_at: When the task was claimed
+- retry_count: Number of retry attempts
+- task_payload: The serialized task dictionary
+"""
+
+from collections.abc import Sequence
+
+import sqlalchemy as sa
+from alembic import context, op
+from sqlalchemy.dialects import postgresql
+
+# revision identifiers, used by Alembic.
+revision: str = "l7g8h9i0j1k2"
+down_revision: str | Sequence[str] | None = "k6f7g8h9i0j1"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (required for multi-tenant support)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    """Add worker columns to async_operations."""
+    schema = _get_schema_prefix()
+
+    # Add worker_id column (ID of worker that claimed the task)
+    op.add_column(
+        "async_operations",
+        sa.Column("worker_id", sa.Text(), nullable=True),
+        schema=context.config.get_main_option("target_schema") or None,
+    )
+
+    # Add claimed_at column (when task was claimed by worker)
+    op.add_column(
+        "async_operations",
+        sa.Column("claimed_at", postgresql.TIMESTAMP(timezone=True), nullable=True),
+        schema=context.config.get_main_option("target_schema") or None,
+    )
+
+    # Add retry_count column (number of retry attempts)
+    op.add_column(
+        "async_operations",
+        sa.Column("retry_count", sa.Integer(), server_default="0", nullable=False),
+        schema=context.config.get_main_option("target_schema") or None,
+    )
+
+    # Add task_payload column (serialized task dictionary)
+    op.add_column(
+        "async_operations",
+        sa.Column(
+            "task_payload",
+            postgresql.JSONB(astext_type=sa.Text()),
+            nullable=True,
+        ),
+        schema=context.config.get_main_option("target_schema") or None,
+    )
+
+    # Add index for efficient worker polling (pending tasks ordered by creation time)
+    op.execute(
+        f"CREATE INDEX idx_async_operations_pending_claim ON {schema}async_operations (status, created_at) "
+        f"WHERE status = 'pending' AND task_payload IS NOT NULL"
+    )
+
+    # Add index for finding tasks by worker_id (for decommissioning)
+    op.execute(
+        f"CREATE INDEX idx_async_operations_worker_id ON {schema}async_operations (worker_id) WHERE worker_id IS NOT NULL"
+    )
+
+
+def downgrade() -> None:
+    """Remove worker columns from async_operations."""
+    schema = _get_schema_prefix()
+
+    # Drop indexes
+    op.execute(f"DROP INDEX IF EXISTS {schema}idx_async_operations_pending_claim")
+    op.execute(f"DROP INDEX IF EXISTS {schema}idx_async_operations_worker_id")
+
+    # Drop columns
+    op.drop_column(
+        "async_operations",
+        "task_payload",
+        schema=context.config.get_main_option("target_schema") or None,
+    )
+    op.drop_column(
+        "async_operations",
+        "retry_count",
+        schema=context.config.get_main_option("target_schema") or None,
+    )
+    op.drop_column(
+        "async_operations",
+        "claimed_at",
+        schema=context.config.get_main_option("target_schema") or None,
+    )
+    op.drop_column(
+        "async_operations",
+        "worker_id",
+        schema=context.config.get_main_option("target_schema") or None,
+    )

hindsight-api/hindsight_api/alembic/versions/m8h9i0j1k2l3_mental_model_id_to_text.py

@@ -0,0 +1,41 @@
+"""mental_model_id_to_text
+
+Revision ID: m8h9i0j1k2l3
+Revises: l7g8h9i0j1k2
+Create Date: 2026-01-19 00:00:00.000000
+
+This migration changes the mental_models.id column from VARCHAR(64) to TEXT
+to support longer model IDs (e.g., entity names that exceed 64 characters).
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+# revision identifiers, used by Alembic.
+revision: str = "m8h9i0j1k2l3"
+down_revision: str | Sequence[str] | None = "l7g8h9i0j1k2"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (required for multi-tenant support)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    """Change mental_models.id from VARCHAR(64) to TEXT."""
+    schema = _get_schema_prefix()
+
+    # Alter the id column type from VARCHAR(64) to TEXT
+    op.execute(f"ALTER TABLE {schema}mental_models ALTER COLUMN id TYPE TEXT")
+
+
+def downgrade() -> None:
+    """Revert mental_models.id from TEXT to VARCHAR(64)."""
+    schema = _get_schema_prefix()
+
+    # Note: This may fail if any id values exceed 64 characters
+    op.execute(f"ALTER TABLE {schema}mental_models ALTER COLUMN id TYPE VARCHAR(64)")

hindsight-api/hindsight_api/alembic/versions/n9i0j1k2l3m4_learnings_and_pinned_reflections.py

@@ -0,0 +1,317 @@
+"""learnings_and_pinned_reflections
+
+Revision ID: n9i0j1k2l3m4
+Revises: m8h9i0j1k2l3
+Create Date: 2026-01-21 00:00:00.000000
+
+This migration:
+1. Creates the 'learnings' table for automatic bottom-up consolidation
+2. Creates the 'pinned_reflections' table for user-curated living documents
+3. Adds consolidation tracking columns to the 'banks' table
+"""
+
+import os
+from collections.abc import Sequence
+
+from alembic import context, op
+from sqlalchemy import text
+
+# revision identifiers, used by Alembic.
+revision: str = "n9i0j1k2l3m4"
+down_revision: str | Sequence[str] | None = "m8h9i0j1k2l3"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (required for multi-tenant support)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def _detect_vector_extension() -> str:
+    """
+    Detect or validate vector extension: 'pgvector', 'vchord', or 'pgvectorscale'.
+    Respects HINDSIGHT_API_VECTOR_EXTENSION env var if set.
+    """
+    conn = op.get_bind()
+    vector_extension = os.getenv("HINDSIGHT_API_VECTOR_EXTENSION", "pgvector").lower()
+
+    # Validate configured extension is installed
+    if vector_extension == "pgvectorscale":
+        # pgvectorscale/DiskANN requires pgvector
+        pgvector_check = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vector'")).scalar()
+        if not pgvector_check:
+            raise RuntimeError(
+                "DiskANN requires pgvector. Install with: CREATE EXTENSION vector; then vectorscale or pg_diskann CASCADE;"
+            )
+        # Check for either vectorscale (open source) or pg_diskann (Azure)
+        vectorscale_check = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vectorscale'")).scalar()
+        pg_diskann_check = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'pg_diskann'")).scalar()
+
+        if vectorscale_check:
+            return "pgvectorscale"
+        elif pg_diskann_check:
+            return "pg_diskann"
+        else:
+            raise RuntimeError(
+                "Configured vector extension 'pgvectorscale' not found. Install either:\n"
+                "  - pgvectorscale: CREATE EXTENSION vectorscale CASCADE;\n"
+                "  - pg_diskann (Azure): CREATE EXTENSION pg_diskann CASCADE;"
+            )
+    elif vector_extension == "vchord":
+        vchord_check = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vchord'")).scalar()
+        if not vchord_check:
+            raise RuntimeError(
+                "Configured vector extension 'vchord' not found. Install it with: CREATE EXTENSION vchord CASCADE;"
+            )
+        return "vchord"
+    elif vector_extension == "pgvector":
+        pgvector_check = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vector'")).scalar()
+        if not pgvector_check:
+            raise RuntimeError(
+                "Configured vector extension 'pgvector' not found. Install it with: CREATE EXTENSION vector;"
+            )
+        return "pgvector"
+    else:
+        raise ValueError(
+            f"Invalid HINDSIGHT_API_VECTOR_EXTENSION: {vector_extension}. Must be 'pgvector', 'vchord', or 'pgvectorscale'"
+        )
+
+
+def _detect_text_search_extension() -> str:
+    """
+    Detect or validate text search extension: 'native', 'vchord', or 'pg_textsearch'.
+    Respects HINDSIGHT_API_TEXT_SEARCH_EXTENSION env var.
+    Creates the extension if needed.
+    """
+    text_search_extension = os.getenv("HINDSIGHT_API_TEXT_SEARCH_EXTENSION", "native").lower()
+
+    if text_search_extension == "vchord":
+        # Create vchord_bm25 extension if not exists
+        try:
+            op.execute("CREATE EXTENSION IF NOT EXISTS vchord_bm25 CASCADE")
+        except Exception:
+            # Extension might already exist or user lacks permissions - verify it exists
+            conn = op.get_bind()
+            result = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vchord_bm25'")).fetchone()
+            if not result:
+                # Extension truly doesn't exist - re-raise the error
+                raise
+        return "vchord"
+    elif text_search_extension == "pg_textsearch":
+        # Create pg_textsearch extension if not exists
+        try:
+            op.execute("CREATE EXTENSION IF NOT EXISTS pg_textsearch CASCADE")
+        except Exception:
+            # Extension might already exist or user lacks permissions - verify it exists
+            conn = op.get_bind()
+            result = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'pg_textsearch'")).fetchone()
+            if not result:
+                # Extension truly doesn't exist - re-raise the error
+                raise
+        return "pg_textsearch"
+    elif text_search_extension == "native":
+        return "native"
+    else:
+        raise ValueError(
+            f"Invalid HINDSIGHT_API_TEXT_SEARCH_EXTENSION: {text_search_extension}. Must be 'native', 'vchord', or 'pg_textsearch'"
+        )
+
+
+def upgrade() -> None:
+    """Create learnings and pinned_reflections tables."""
+    schema = _get_schema_prefix()
+
+    # Detect which vector extension is available
+    vector_ext = _detect_vector_extension()
+
+    # Detect which text search extension to use
+    text_search_ext = _detect_text_search_extension()
+
+    # 1. Create learnings table
+    op.execute(f"""
+        CREATE TABLE {schema}learnings (
+            id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+            bank_id VARCHAR(64) NOT NULL,
+            text TEXT NOT NULL,
+            proof_count INT NOT NULL DEFAULT 1,
+            history JSONB DEFAULT '[]'::jsonb,
+            mission_context VARCHAR(64),
+            pre_mission_change BOOLEAN DEFAULT FALSE,
+            embedding vector(384),
+            tags VARCHAR[] DEFAULT ARRAY[]::VARCHAR[],
+            created_at TIMESTAMP WITH TIME ZONE NOT NULL DEFAULT now(),
+            updated_at TIMESTAMP WITH TIME ZONE NOT NULL DEFAULT now()
+        )
+    """)
+
+    # Add foreign key constraint
+    op.execute(f"""
+        ALTER TABLE {schema}learnings
+        ADD CONSTRAINT fk_learnings_bank_id
+        FOREIGN KEY (bank_id) REFERENCES {schema}banks(bank_id) ON DELETE CASCADE
+    """)
+
+    # Indexes for learnings
+    op.execute(f"CREATE INDEX idx_learnings_bank_id ON {schema}learnings(bank_id)")
+
+    # Create vector index based on detected extension
+    if vector_ext == "pgvectorscale":
+        op.execute(f"""
+            CREATE INDEX idx_learnings_embedding ON {schema}learnings
+            USING diskann (embedding vector_cosine_ops)
+            WITH (num_neighbors = 50)
+        """)
+    elif vector_ext == "pg_diskann":
+        op.execute(f"""
+            CREATE INDEX idx_learnings_embedding ON {schema}learnings
+            USING diskann (embedding vector_cosine_ops)
+            WITH (max_neighbors = 50)
+        """)
+    elif vector_ext == "vchord":
+        op.execute(f"""
+            CREATE INDEX idx_learnings_embedding ON {schema}learnings
+            USING vchordrq (embedding vector_l2_ops)
+        """)
+    else:  # pgvector
+        op.execute(f"""
+            CREATE INDEX idx_learnings_embedding ON {schema}learnings
+            USING hnsw (embedding vector_cosine_ops)
+        """)
+
+    op.execute(f"CREATE INDEX idx_learnings_tags ON {schema}learnings USING GIN(tags)")
+
+    # Full-text search for learnings
+    if text_search_ext == "vchord":
+        # VectorChord BM25: bm25vector type (no GENERATED - tokenization happens on INSERT)
+        # Note: vchord_bm25 extension creates types in bm25_catalog schema
+        op.execute(f"""
+            ALTER TABLE {schema}learnings ADD COLUMN search_vector bm25_catalog.bm25vector
+        """)
+        op.execute(f"""
+            CREATE INDEX idx_learnings_text_search ON {schema}learnings
+            USING bm25 (search_vector bm25_catalog.bm25_ops)
+        """)
+    elif text_search_ext == "pg_textsearch":
+        # Timescale pg_textsearch: dummy TEXT column for consistency (indexes operate on base columns directly)
+        op.execute(f"""
+            ALTER TABLE {schema}learnings ADD COLUMN search_vector TEXT
+        """)
+        op.execute(f"""
+            CREATE INDEX idx_learnings_text_search ON {schema}learnings
+            USING bm25(text) WITH (text_config='english')
+        """)
+    else:  # native
+        # Native PostgreSQL: tsvector with automatic generation
+        op.execute(f"""
+            ALTER TABLE {schema}learnings ADD COLUMN search_vector tsvector
+            GENERATED ALWAYS AS (to_tsvector('english', text)) STORED
+        """)
+        op.execute(f"CREATE INDEX idx_learnings_text_search ON {schema}learnings USING gin(search_vector)")
+
+    # 2. Create pinned_reflections table
+    op.execute(f"""
+        CREATE TABLE {schema}pinned_reflections (
+            id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+            bank_id VARCHAR(64) NOT NULL,
+            name VARCHAR(256) NOT NULL,
+            source_query TEXT NOT NULL,
+            content TEXT NOT NULL,
+            embedding vector(384),
+            tags VARCHAR[] DEFAULT ARRAY[]::VARCHAR[],
+            last_refreshed_at TIMESTAMP WITH TIME ZONE NOT NULL DEFAULT now(),
+            created_at TIMESTAMP WITH TIME ZONE NOT NULL DEFAULT now()
+        )
+    """)
+
+    # Add foreign key constraint
+    op.execute(f"""
+        ALTER TABLE {schema}pinned_reflections
+        ADD CONSTRAINT fk_pinned_reflections_bank_id
+        FOREIGN KEY (bank_id) REFERENCES {schema}banks(bank_id) ON DELETE CASCADE
+    """)
+
+    # Indexes for pinned_reflections
+    op.execute(f"CREATE INDEX idx_pinned_reflections_bank_id ON {schema}pinned_reflections(bank_id)")
+
+    # Create vector index based on detected extension
+    if vector_ext == "pgvectorscale":
+        op.execute(f"""
+            CREATE INDEX idx_pinned_reflections_embedding ON {schema}pinned_reflections
+            USING diskann (embedding vector_cosine_ops)
+            WITH (num_neighbors = 50)
+        """)
+    elif vector_ext == "pg_diskann":
+        op.execute(f"""
+            CREATE INDEX idx_pinned_reflections_embedding ON {schema}pinned_reflections
+            USING diskann (embedding vector_cosine_ops)
+            WITH (max_neighbors = 50)
+        """)
+    elif vector_ext == "vchord":
+        op.execute(f"""
+            CREATE INDEX idx_pinned_reflections_embedding ON {schema}pinned_reflections
+            USING vchordrq (embedding vector_l2_ops)
+        """)
+    else:  # pgvector
+        op.execute(f"""
+            CREATE INDEX idx_pinned_reflections_embedding ON {schema}pinned_reflections
+            USING hnsw (embedding vector_cosine_ops)
+        """)
+
+    op.execute(f"CREATE INDEX idx_pinned_reflections_tags ON {schema}pinned_reflections USING GIN(tags)")
+
+    # Full-text search for pinned_reflections
+    if text_search_ext == "vchord":
+        # VectorChord BM25: bm25vector type (no GENERATED - tokenization happens on INSERT/UPDATE)
+        # Note: vchord_bm25 extension creates types in bm25_catalog schema
+        op.execute(f"""
+            ALTER TABLE {schema}pinned_reflections ADD COLUMN search_vector bm25_catalog.bm25vector
+        """)
+        op.execute(f"""
+            CREATE INDEX idx_pinned_reflections_text_search ON {schema}pinned_reflections
+            USING bm25 (search_vector bm25_catalog.bm25_ops)
+        """)
+    elif text_search_ext == "pg_textsearch":
+        # Timescale pg_textsearch: dummy TEXT column for consistency (indexes operate on base columns directly)
+        op.execute(f"""
+            ALTER TABLE {schema}pinned_reflections ADD COLUMN search_vector TEXT
+        """)
+        op.execute(f"""
+            CREATE INDEX idx_pinned_reflections_text_search ON {schema}pinned_reflections
+            USING bm25(content)
+            WITH (text_config='english')
+        """)
+    else:  # native
+        # Native PostgreSQL: tsvector with automatic generation
+        op.execute(f"""
+            ALTER TABLE {schema}pinned_reflections ADD COLUMN search_vector tsvector
+            GENERATED ALWAYS AS (to_tsvector('english', COALESCE(name, '') || ' ' || content)) STORED
+        """)
+        op.execute(f"""
+            CREATE INDEX idx_pinned_reflections_text_search ON {schema}pinned_reflections
+            USING gin(search_vector)
+        """)
+
+    # 3. Add consolidation tracking columns to banks table
+    op.execute(f"""
+        ALTER TABLE {schema}banks
+        ADD COLUMN IF NOT EXISTS last_consolidated_at TIMESTAMP WITH TIME ZONE
+    """)
+    op.execute(f"""
+        ALTER TABLE {schema}banks
+        ADD COLUMN IF NOT EXISTS mission_changed_at TIMESTAMP WITH TIME ZONE
+    """)
+
+
+def downgrade() -> None:
+    """Drop learnings and pinned_reflections tables."""
+    schema = _get_schema_prefix()
+
+    # Drop tables
+    op.execute(f"DROP TABLE IF EXISTS {schema}learnings CASCADE")
+    op.execute(f"DROP TABLE IF EXISTS {schema}pinned_reflections CASCADE")
+
+    # Remove columns from banks
+    op.execute(f"ALTER TABLE {schema}banks DROP COLUMN IF EXISTS last_consolidated_at")
+    op.execute(f"ALTER TABLE {schema}banks DROP COLUMN IF EXISTS mission_changed_at")

hindsight-api/hindsight_api/alembic/versions/o0j1k2l3m4n5_migrate_mental_models_data.py

@@ -0,0 +1,113 @@
+"""migrate_mental_models_data
+
+Revision ID: o0j1k2l3m4n5
+Revises: n9i0j1k2l3m4
+Create Date: 2026-01-21 00:00:00.000000
+
+This migration:
+1. Migrates existing 'pinned' mental models to the new 'pinned_reflections' table
+2. Migrates existing 'learned' mental models to the new 'learnings' table
+3. Deletes non-directive mental models (structural, emergent, pinned, learned)
+4. Drops the mental_model_versions table (no longer used)
+5. Adds a CHECK constraint that only 'directive' subtype is allowed
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+# revision identifiers, used by Alembic.
+revision: str = "o0j1k2l3m4n5"
+down_revision: str | Sequence[str] | None = "n9i0j1k2l3m4"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (required for multi-tenant support)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    """Migrate data and clean up old mental models."""
+    schema = _get_schema_prefix()
+
+    # 1. Migrate 'pinned' mental models to pinned_reflections
+    # For pinned models, the first observation's content becomes the pinned reflection content
+    op.execute(f"""
+        INSERT INTO {schema}pinned_reflections (bank_id, name, source_query, content, tags, created_at)
+        SELECT
+            bank_id,
+            name,
+            description AS source_query,
+            COALESCE(
+                observations->'observations'->0->>'content',
+                description,
+                ''
+            ) AS content,
+            tags,
+            created_at
+        FROM {schema}mental_models
+        WHERE subtype = 'pinned'
+        ON CONFLICT DO NOTHING
+    """)
+
+    # 2. Migrate 'learned' mental models to learnings
+    # Each observation in a learned model becomes a separate learning
+    op.execute(f"""
+        INSERT INTO {schema}learnings (bank_id, text, proof_count, tags, created_at)
+        SELECT
+            mm.bank_id,
+            obs->>'content' AS text,
+            GREATEST(1, COALESCE(jsonb_array_length(obs->'evidence'), 1)) AS proof_count,
+            mm.tags,
+            mm.created_at
+        FROM {schema}mental_models mm,
+        LATERAL jsonb_array_elements(mm.observations->'observations') AS obs
+        WHERE mm.subtype = 'learned'
+          AND obs->>'content' IS NOT NULL
+          AND obs->>'content' != ''
+        ON CONFLICT DO NOTHING
+    """)
+
+    # 3. Delete all non-directive mental models (they've been migrated or are obsolete)
+    op.execute(f"""
+        DELETE FROM {schema}mental_models
+        WHERE subtype != 'directive'
+    """)
+
+    # 4. Drop the mental_model_versions table (no longer used)
+    op.execute(f"DROP TABLE IF EXISTS {schema}mental_model_versions CASCADE")
+
+    # 5. Drop old constraints and add new one that only allows 'directive'
+    op.execute(f"ALTER TABLE {schema}mental_models DROP CONSTRAINT IF EXISTS ck_mental_models_subtype")
+    op.execute(f"""
+        ALTER TABLE {schema}mental_models
+        ADD CONSTRAINT ck_mental_models_subtype CHECK (subtype = 'directive')
+    """)
+
+
+def downgrade() -> None:
+    """Reverse the migration (data migration is one-way, so this just removes constraints)."""
+    schema = _get_schema_prefix()
+
+    # Remove the directive-only constraint
+    op.execute(f"ALTER TABLE {schema}mental_models DROP CONSTRAINT IF EXISTS ck_mental_models_subtype")
+
+    # Re-create mental_model_versions table
+    op.execute(f"""
+        CREATE TABLE IF NOT EXISTS {schema}mental_model_versions (
+            id SERIAL PRIMARY KEY,
+            bank_id VARCHAR(64) NOT NULL,
+            model_id VARCHAR(128) NOT NULL,
+            version INT NOT NULL,
+            observations JSONB NOT NULL,
+            created_at TIMESTAMP WITH TIME ZONE NOT NULL DEFAULT now()
+        )
+    """)
+    op.execute(
+        f"CREATE INDEX IF NOT EXISTS idx_mm_versions_lookup ON {schema}mental_model_versions(bank_id, model_id, version DESC)"
+    )
+
+    # Note: Data migration cannot be reversed - pinned_reflections and learnings data remains

hindsight-api/hindsight_api/alembic/versions/p1k2l3m4n5o6_new_knowledge_architecture.py

@@ -0,0 +1,194 @@
+"""new_knowledge_architecture
+
+Revision ID: p1k2l3m4n5o6
+Revises: o0j1k2l3m4n5
+Create Date: 2026-01-21 00:00:00.000000
+
+This migration implements the new knowledge architecture:
+1. Drops the 'learnings' table (mental models are now in memory_units)
+2. Renames 'pinned_reflections' to 'reflections'
+3. Drops the 'mental_models' table completely
+4. Creates 'directives' table for hard rules
+5. Adds mental model support columns to 'memory_units' (proof_count, source_memory_ids, history)
+
+The new architecture:
+- Directives: Hard rules in their own table
+- Mental Models: Stored in memory_units with fact_type='mental_model'
+- Reflections: User-curated documents (renamed from pinned_reflections)
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+# revision identifiers, used by Alembic.
+revision: str = "p1k2l3m4n5o6"
+down_revision: str | Sequence[str] | None = "o0j1k2l3m4n5"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (required for multi-tenant support)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    """Implement new knowledge architecture."""
+    schema = _get_schema_prefix()
+
+    # 1. Drop the learnings table (mental models will be in memory_units)
+    op.execute(f"DROP TABLE IF EXISTS {schema}learnings CASCADE")
+
+    # 2. Rename pinned_reflections to reflections
+    op.execute(f"ALTER TABLE IF EXISTS {schema}pinned_reflections RENAME TO reflections")
+
+    # Rename indexes for reflections
+    op.execute(f"ALTER INDEX IF EXISTS {schema}idx_pinned_reflections_bank_id RENAME TO idx_reflections_bank_id")
+    op.execute(f"ALTER INDEX IF EXISTS {schema}idx_pinned_reflections_embedding RENAME TO idx_reflections_embedding")
+    op.execute(f"ALTER INDEX IF EXISTS {schema}idx_pinned_reflections_tags RENAME TO idx_reflections_tags")
+    op.execute(
+        f"ALTER INDEX IF EXISTS {schema}idx_pinned_reflections_text_search RENAME TO idx_reflections_text_search"
+    )
+
+    # Rename foreign key constraint
+    op.execute(f"""
+        ALTER TABLE {schema}reflections
+        DROP CONSTRAINT IF EXISTS fk_pinned_reflections_bank_id
+    """)
+    op.execute(f"""
+        ALTER TABLE {schema}reflections
+        ADD CONSTRAINT fk_reflections_bank_id
+        FOREIGN KEY (bank_id) REFERENCES {schema}banks(bank_id) ON DELETE CASCADE
+    """)
+
+    # 3. Drop the mental_models table completely
+    op.execute(f"DROP TABLE IF EXISTS {schema}mental_models CASCADE")
+
+    # 4. Create directives table
+    op.execute(f"""
+        CREATE TABLE {schema}directives (
+            id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+            bank_id VARCHAR(64) NOT NULL,
+            name VARCHAR(256) NOT NULL,
+            content TEXT NOT NULL,
+            priority INT NOT NULL DEFAULT 0,
+            is_active BOOLEAN NOT NULL DEFAULT TRUE,
+            tags VARCHAR[] DEFAULT ARRAY[]::VARCHAR[],
+            created_at TIMESTAMP WITH TIME ZONE NOT NULL DEFAULT now(),
+            updated_at TIMESTAMP WITH TIME ZONE NOT NULL DEFAULT now()
+        )
+    """)
+
+    # Add foreign key and indexes for directives
+    op.execute(f"""
+        ALTER TABLE {schema}directives
+        ADD CONSTRAINT fk_directives_bank_id
+        FOREIGN KEY (bank_id) REFERENCES {schema}banks(bank_id) ON DELETE CASCADE
+    """)
+    op.execute(f"CREATE INDEX idx_directives_bank_id ON {schema}directives(bank_id)")
+    op.execute(f"CREATE INDEX idx_directives_bank_active ON {schema}directives(bank_id, is_active)")
+    op.execute(f"CREATE INDEX idx_directives_tags ON {schema}directives USING GIN(tags)")
+
+    # 5. Add mental model support columns to memory_units
+    # proof_count: Number of memories that support this mental model
+    op.execute(f"""
+        ALTER TABLE {schema}memory_units
+        ADD COLUMN IF NOT EXISTS proof_count INT DEFAULT 1
+    """)
+
+    # source_memory_ids: Array of memory IDs that consolidated into this mental model
+    op.execute(f"""
+        ALTER TABLE {schema}memory_units
+        ADD COLUMN IF NOT EXISTS source_memory_ids UUID[] DEFAULT ARRAY[]::UUID[]
+    """)
+
+    # history: JSONB array tracking changes to mental models
+    op.execute(f"""
+        ALTER TABLE {schema}memory_units
+        ADD COLUMN IF NOT EXISTS history JSONB DEFAULT '[]'::jsonb
+    """)
+
+    # Add index for finding mental models
+    op.execute(f"""
+        CREATE INDEX IF NOT EXISTS idx_memory_units_mental_models
+        ON {schema}memory_units(bank_id, fact_type)
+        WHERE fact_type = 'mental_model'
+    """)
+
+    # 6. Update fact_type check constraint to include 'mental_model'
+    op.execute(f"ALTER TABLE {schema}memory_units DROP CONSTRAINT IF EXISTS memory_units_fact_type_check")
+    op.execute(f"""
+        ALTER TABLE {schema}memory_units
+        ADD CONSTRAINT memory_units_fact_type_check
+        CHECK (fact_type IN ('world', 'experience', 'opinion', 'observation', 'mental_model'))
+    """)
+
+
+def downgrade() -> None:
+    """Reverse the migration."""
+    schema = _get_schema_prefix()
+
+    # Restore original fact_type check constraint (without 'mental_model')
+    op.execute(f"ALTER TABLE {schema}memory_units DROP CONSTRAINT IF EXISTS memory_units_fact_type_check")
+    op.execute(f"""
+        ALTER TABLE {schema}memory_units
+        ADD CONSTRAINT memory_units_fact_type_check
+        CHECK (fact_type IN ('world', 'experience', 'opinion', 'observation'))
+    """)
+
+    # Drop mental model columns from memory_units
+    op.execute(f"ALTER TABLE {schema}memory_units DROP COLUMN IF EXISTS proof_count")
+    op.execute(f"ALTER TABLE {schema}memory_units DROP COLUMN IF EXISTS source_memory_ids")
+    op.execute(f"ALTER TABLE {schema}memory_units DROP COLUMN IF EXISTS history")
+    op.execute(f"DROP INDEX IF EXISTS {schema}idx_memory_units_mental_models")
+
+    # Drop directives table
+    op.execute(f"DROP TABLE IF EXISTS {schema}directives CASCADE")
+
+    # Rename reflections back to pinned_reflections
+    op.execute(f"ALTER TABLE IF EXISTS {schema}reflections RENAME TO pinned_reflections")
+
+    # Restore indexes
+    op.execute(f"ALTER INDEX IF EXISTS {schema}idx_reflections_bank_id RENAME TO idx_pinned_reflections_bank_id")
+    op.execute(f"ALTER INDEX IF EXISTS {schema}idx_reflections_embedding RENAME TO idx_pinned_reflections_embedding")
+    op.execute(f"ALTER INDEX IF EXISTS {schema}idx_reflections_tags RENAME TO idx_pinned_reflections_tags")
+    op.execute(
+        f"ALTER INDEX IF EXISTS {schema}idx_reflections_text_search RENAME TO idx_pinned_reflections_text_search"
+    )
+
+    # Restore foreign key
+    op.execute(f"""
+        ALTER TABLE {schema}pinned_reflections
+        DROP CONSTRAINT IF EXISTS fk_reflections_bank_id
+    """)
+    op.execute(f"""
+        ALTER TABLE {schema}pinned_reflections
+        ADD CONSTRAINT fk_pinned_reflections_bank_id
+        FOREIGN KEY (bank_id) REFERENCES {schema}banks(bank_id) ON DELETE CASCADE
+    """)
+
+    # Re-create learnings table
+    op.execute(f"""
+        CREATE TABLE IF NOT EXISTS {schema}learnings (
+            id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+            bank_id VARCHAR(64) NOT NULL,
+            text TEXT NOT NULL,
+            proof_count INT NOT NULL DEFAULT 1,
+            history JSONB DEFAULT '[]'::jsonb,
+            mission_context VARCHAR(64),
+            pre_mission_change BOOLEAN DEFAULT FALSE,
+            embedding vector(384),
+            tags VARCHAR[] DEFAULT ARRAY[]::VARCHAR[],
+            created_at TIMESTAMP WITH TIME ZONE NOT NULL DEFAULT now(),
+            updated_at TIMESTAMP WITH TIME ZONE NOT NULL DEFAULT now()
+        )
+    """)
+    op.execute(f"""
+        ALTER TABLE {schema}learnings
+        ADD CONSTRAINT fk_learnings_bank_id
+        FOREIGN KEY (bank_id) REFERENCES {schema}banks(bank_id) ON DELETE CASCADE
+    """)
+
+    # Note: mental_models table recreation is complex and would need separate handling

hindsight-api/hindsight_api/alembic/versions/q2l3m4n5o6p7_fix_mental_model_fact_type.py

@@ -0,0 +1,50 @@
+"""fix_mental_model_fact_type
+
+Revision ID: q2l3m4n5o6p7
+Revises: p1k2l3m4n5o6
+Create Date: 2026-01-21 13:30:00.000000
+
+Fix the fact_type check constraint to include 'mental_model'.
+This is a fix for p1k2l3m4n5o6 which should have included this change.
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+# revision identifiers, used by Alembic.
+revision: str = "q2l3m4n5o6p7"
+down_revision: str | Sequence[str] | None = "p1k2l3m4n5o6"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (required for multi-tenant support)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    """Add 'mental_model' to the fact_type check constraint."""
+    schema = _get_schema_prefix()
+
+    # Drop the old constraint and add the new one with mental_model included
+    op.execute(f"ALTER TABLE {schema}memory_units DROP CONSTRAINT IF EXISTS memory_units_fact_type_check")
+    op.execute(f"""
+        ALTER TABLE {schema}memory_units
+        ADD CONSTRAINT memory_units_fact_type_check
+        CHECK (fact_type IN ('world', 'experience', 'opinion', 'observation', 'mental_model'))
+    """)
+
+
+def downgrade() -> None:
+    """Remove 'mental_model' from the fact_type check constraint."""
+    schema = _get_schema_prefix()
+
+    op.execute(f"ALTER TABLE {schema}memory_units DROP CONSTRAINT IF EXISTS memory_units_fact_type_check")
+    op.execute(f"""
+        ALTER TABLE {schema}memory_units
+        ADD CONSTRAINT memory_units_fact_type_check
+        CHECK (fact_type IN ('world', 'experience', 'opinion', 'observation'))
+    """)

hindsight-api/hindsight_api/alembic/versions/r3m4n5o6p7q8_add_reflect_response_to_reflections.py

@@ -0,0 +1,47 @@
+"""Add reflect_response JSONB column to reflections
+
+Revision ID: r3m4n5o6p7q8
+Revises: q2l3m4n5o6p7
+Create Date: 2026-01-21
+
+This migration adds a reflect_response JSONB column to store the full
+reflect API response payload, including based_on facts and trace data.
+
+Note: Table was renamed from pinned_reflections to reflections in p1k2l3m4n5o6.
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+revision: str = "r3m4n5o6p7q8"
+down_revision: str | Sequence[str] | None = "q2l3m4n5o6p7"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (required for multi-tenant support)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    """Add reflect_response JSONB column to reflections."""
+    schema = _get_schema_prefix()
+
+    # Add reflect_response column to store the full reflect API response
+    op.execute(f"""
+        ALTER TABLE {schema}reflections
+        ADD COLUMN IF NOT EXISTS reflect_response JSONB
+    """)
+
+
+def downgrade() -> None:
+    """Remove reflect_response column from reflections."""
+    schema = _get_schema_prefix()
+
+    op.execute(f"""
+        ALTER TABLE {schema}reflections
+        DROP COLUMN IF EXISTS reflect_response
+    """)

hindsight-api/hindsight_api/alembic/versions/s4n5o6p7q8r9_add_consolidated_at_to_memory_units.py

@@ -0,0 +1,53 @@
+"""Add consolidated_at column to memory_units for incremental consolidation tracking.
+
+This allows consolidation to track progress at the memory level rather than
+using a bank-level watermark. If consolidation crashes, already-processed
+memories won't be reprocessed.
+
+Revision ID: s4n5o6p7q8r9
+Revises: r3m4n5o6p7q8
+Create Date: 2025-01-22
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+revision: str = "s4n5o6p7q8r9"
+down_revision: str | Sequence[str] | None = "r3m4n5o6p7q8"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (required for multi-tenant support)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    schema = _get_schema_prefix()
+
+    # Add consolidated_at column to memory_units
+    op.execute(
+        f"""
+        ALTER TABLE {schema}memory_units
+        ADD COLUMN IF NOT EXISTS consolidated_at TIMESTAMPTZ DEFAULT NULL
+        """
+    )
+
+    # Create index for efficient querying of unconsolidated memories
+    op.execute(
+        f"""
+        CREATE INDEX IF NOT EXISTS idx_memory_units_unconsolidated
+        ON {schema}memory_units (bank_id, created_at)
+        WHERE consolidated_at IS NULL AND fact_type IN ('experience', 'world')
+        """
+    )
+
+
+def downgrade() -> None:
+    schema = _get_schema_prefix()
+
+    op.execute(f"DROP INDEX IF EXISTS {schema}idx_memory_units_unconsolidated")
+    op.execute(f"ALTER TABLE {schema}memory_units DROP COLUMN IF EXISTS consolidated_at")

hindsight-api/hindsight_api/alembic/versions/t5o6p7q8r9s0_rename_mental_models_to_observations.py

@@ -0,0 +1,134 @@
+"""Rename mental_model fact_type to observation and reflections table to mental_models
+
+Revision ID: t5o6p7q8r9s0
+Revises: s4n5o6p7q8r9
+Create Date: 2026-01-26
+
+This migration implements the terminology rename:
+1. mental_model (fact_type in memory_units) -> observation
+2. reflections table -> mental_models table
+
+The new terminology:
+- Observations: Consolidated knowledge synthesized from facts (was mental_model)
+- Mental Models: Stored reflect responses (was reflections)
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+revision: str = "t5o6p7q8r9s0"
+down_revision: str | Sequence[str] | None = "s4n5o6p7q8r9"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (required for multi-tenant support)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    """Rename mental_model -> observation and reflections -> mental_models."""
+    schema = _get_schema_prefix()
+
+    # 1. Update fact_type values: mental_model -> observation
+    op.execute(f"""
+        UPDATE {schema}memory_units
+        SET fact_type = 'observation'
+        WHERE fact_type = 'mental_model'
+    """)
+
+    # 2. Update the CHECK constraint - remove mental_model, keep observation
+    op.execute(f"ALTER TABLE {schema}memory_units DROP CONSTRAINT IF EXISTS memory_units_fact_type_check")
+    op.execute(f"""
+        ALTER TABLE {schema}memory_units
+        ADD CONSTRAINT memory_units_fact_type_check
+        CHECK (fact_type IN ('world', 'experience', 'opinion', 'observation'))
+    """)
+
+    # 3. Rename the index for observations (was for mental_models)
+    op.execute(f"DROP INDEX IF EXISTS {schema}idx_memory_units_mental_models")
+    op.execute(f"""
+        CREATE INDEX IF NOT EXISTS idx_memory_units_observations
+        ON {schema}memory_units(bank_id, fact_type)
+        WHERE fact_type = 'observation'
+    """)
+
+    # 4. Update the unconsolidated index to not filter by fact_type since observations
+    # are now the consolidated type
+    op.execute(f"DROP INDEX IF EXISTS {schema}idx_memory_units_unconsolidated")
+    op.execute(f"""
+        CREATE INDEX IF NOT EXISTS idx_memory_units_unconsolidated
+        ON {schema}memory_units (bank_id, created_at)
+        WHERE consolidated_at IS NULL AND fact_type IN ('experience', 'world')
+    """)
+
+    # 5. Rename reflections table to mental_models
+    op.execute(f"ALTER TABLE IF EXISTS {schema}reflections RENAME TO mental_models")
+
+    # 6. Rename indexes for mental_models (was reflections)
+    op.execute(f"ALTER INDEX IF EXISTS {schema}idx_reflections_bank_id RENAME TO idx_mental_models_bank_id")
+    op.execute(f"ALTER INDEX IF EXISTS {schema}idx_reflections_embedding RENAME TO idx_mental_models_embedding")
+    op.execute(f"ALTER INDEX IF EXISTS {schema}idx_reflections_tags RENAME TO idx_mental_models_tags")
+    op.execute(f"ALTER INDEX IF EXISTS {schema}idx_reflections_text_search RENAME TO idx_mental_models_text_search")
+
+    # 7. Rename foreign key constraint
+    op.execute(f"""
+        ALTER TABLE {schema}mental_models
+        DROP CONSTRAINT IF EXISTS fk_reflections_bank_id
+    """)
+    op.execute(f"""
+        ALTER TABLE {schema}mental_models
+        ADD CONSTRAINT fk_mental_models_bank_id
+        FOREIGN KEY (bank_id) REFERENCES {schema}banks(bank_id) ON DELETE CASCADE
+    """)
+
+
+def downgrade() -> None:
+    """Reverse: observation -> mental_model and mental_models -> reflections."""
+    schema = _get_schema_prefix()
+
+    # 1. Rename mental_models table back to reflections
+    op.execute(f"ALTER TABLE IF EXISTS {schema}mental_models RENAME TO reflections")
+
+    # 2. Rename indexes back
+    op.execute(f"ALTER INDEX IF EXISTS {schema}idx_mental_models_bank_id RENAME TO idx_reflections_bank_id")
+    op.execute(f"ALTER INDEX IF EXISTS {schema}idx_mental_models_embedding RENAME TO idx_reflections_embedding")
+    op.execute(f"ALTER INDEX IF EXISTS {schema}idx_mental_models_tags RENAME TO idx_reflections_tags")
+    op.execute(f"ALTER INDEX IF EXISTS {schema}idx_mental_models_text_search RENAME TO idx_reflections_text_search")
+
+    # 3. Rename foreign key back
+    op.execute(f"""
+        ALTER TABLE {schema}reflections
+        DROP CONSTRAINT IF EXISTS fk_mental_models_bank_id
+    """)
+    op.execute(f"""
+        ALTER TABLE {schema}reflections
+        ADD CONSTRAINT fk_reflections_bank_id
+        FOREIGN KEY (bank_id) REFERENCES {schema}banks(bank_id) ON DELETE CASCADE
+    """)
+
+    # 4. Update fact_type values: observation -> mental_model
+    op.execute(f"""
+        UPDATE {schema}memory_units
+        SET fact_type = 'mental_model'
+        WHERE fact_type = 'observation'
+    """)
+
+    # 5. Update the CHECK constraint back
+    op.execute(f"ALTER TABLE {schema}memory_units DROP CONSTRAINT IF EXISTS memory_units_fact_type_check")
+    op.execute(f"""
+        ALTER TABLE {schema}memory_units
+        ADD CONSTRAINT memory_units_fact_type_check
+        CHECK (fact_type IN ('world', 'experience', 'opinion', 'observation', 'mental_model'))
+    """)
+
+    # 6. Rename index back
+    op.execute(f"DROP INDEX IF EXISTS {schema}idx_memory_units_observations")
+    op.execute(f"""
+        CREATE INDEX IF NOT EXISTS idx_memory_units_mental_models
+        ON {schema}memory_units(bank_id, fact_type)
+        WHERE fact_type = 'mental_model'
+    """)

hindsight-api/hindsight_api/alembic/versions/u6p7q8r9s0t1_mental_models_text_id.py

@@ -0,0 +1,41 @@
+"""Change mental_models.id from UUID to TEXT
+
+Revision ID: u6p7q8r9s0t1
+Revises: t5o6p7q8r9s0
+Create Date: 2026-01-27
+
+This migration changes the mental_models.id column from UUID to TEXT
+to support user-defined text identifiers like 'team-communication' instead of UUIDs.
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+revision: str = "u6p7q8r9s0t1"
+down_revision: str | Sequence[str] | None = "t5o6p7q8r9s0"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (required for multi-tenant support)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    """Change mental_models.id from UUID to TEXT."""
+    schema = _get_schema_prefix()
+
+    # Change the id column type from UUID to TEXT
+    # Existing UUIDs will be converted to their string representation
+    op.execute(f"ALTER TABLE {schema}mental_models ALTER COLUMN id TYPE TEXT USING id::TEXT")
+
+
+def downgrade() -> None:
+    """Revert mental_models.id from TEXT to UUID."""
+    schema = _get_schema_prefix()
+
+    # Note: This will fail if any id values are not valid UUIDs
+    op.execute(f"ALTER TABLE {schema}mental_models ALTER COLUMN id TYPE UUID USING id::UUID")

hindsight-api/hindsight_api/alembic/versions/v7q8r9s0t1u2_add_max_tokens_to_mental_models.py

@@ -0,0 +1,50 @@
+"""Add max_tokens and trigger columns to mental_models
+
+Revision ID: v7q8r9s0t1u2
+Revises: u6p7q8r9s0t1
+Create Date: 2026-01-27
+
+This migration adds:
+- max_tokens column: token limit for content generation during refresh
+- trigger column: JSONB for trigger settings (e.g., refresh_after_consolidation)
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+revision: str = "v7q8r9s0t1u2"
+down_revision: str | Sequence[str] | None = "u6p7q8r9s0t1"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (required for multi-tenant support)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    """Add max_tokens and trigger columns to mental_models."""
+    schema = _get_schema_prefix()
+
+    op.execute(f"""
+        ALTER TABLE {schema}mental_models
+        ADD COLUMN IF NOT EXISTS max_tokens INT NOT NULL DEFAULT 2048
+    """)
+
+    # trigger column stores trigger settings as JSONB
+    # Default: refresh_after_consolidation = false (not "real time")
+    op.execute(f"""
+        ALTER TABLE {schema}mental_models
+        ADD COLUMN IF NOT EXISTS trigger JSONB NOT NULL DEFAULT '{{"refresh_after_consolidation": false}}'::jsonb
+    """)
+
+
+def downgrade() -> None:
+    """Remove max_tokens and trigger columns from mental_models."""
+    schema = _get_schema_prefix()
+
+    op.execute(f"ALTER TABLE {schema}mental_models DROP COLUMN IF EXISTS max_tokens")
+    op.execute(f"ALTER TABLE {schema}mental_models DROP COLUMN IF EXISTS trigger")

hindsight-api/hindsight_api/alembic/versions/w8r9s0t1u2v3_fix_mental_models_pk_isolation.py

@@ -0,0 +1,60 @@
+"""Fix mental_models primary key to be scoped per bank
+
+Revision ID: w8r9s0t1u2v3
+Revises: v7q8r9s0t1u2
+Create Date: 2026-02-05
+
+This migration fixes a critical bank isolation bug where mental_models.id was
+globally unique across all banks instead of being scoped per bank. This caused
+conflicts when different banks tried to use the same custom ID.
+
+CRITICAL FIX: Changes primary key from (id) to (bank_id, id) to ensure proper isolation.
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+revision: str = "w8r9s0t1u2v3"
+down_revision: str | Sequence[str] | None = "v7q8r9s0t1u2"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (required for multi-tenant support)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    """Change mental_models primary key from (id) to (bank_id, id) for proper bank isolation."""
+    schema = _get_schema_prefix()
+
+    # Drop the old primary key constraint (just id)
+    # Note: The constraint might be named differently on different DBs
+    # Try both old names (pinned_reflections_pkey from original, mental_models_pkey from rename)
+    op.execute(f"ALTER TABLE {schema}mental_models DROP CONSTRAINT IF EXISTS pinned_reflections_pkey")
+    op.execute(f"ALTER TABLE {schema}mental_models DROP CONSTRAINT IF EXISTS mental_models_pkey")
+
+    # Create the new composite primary key (bank_id, id)
+    # This ensures IDs are scoped per bank, not globally
+    op.execute(f"""
+        ALTER TABLE {schema}mental_models
+        ADD CONSTRAINT mental_models_pkey PRIMARY KEY (bank_id, id)
+    """)
+
+
+def downgrade() -> None:
+    """Revert mental_models primary key from (bank_id, id) to (id)."""
+    schema = _get_schema_prefix()
+
+    # Drop the composite primary key
+    op.execute(f"ALTER TABLE {schema}mental_models DROP CONSTRAINT IF EXISTS mental_models_pkey")
+
+    # Restore the old primary key (just id)
+    # WARNING: This downgrade will fail if there are duplicate IDs across banks
+    op.execute(f"""
+        ALTER TABLE {schema}mental_models
+        ADD CONSTRAINT mental_models_pkey PRIMARY KEY (id)
+    """)

hindsight-api/hindsight_api/alembic/versions/x9s0t1u2v3w4_add_bank_config_column.py

@@ -0,0 +1,64 @@
+"""Add config JSONB column to banks table for hierarchical configuration
+
+Revision ID: x9s0t1u2v3w4
+Revises: w8r9s0t1u2v3
+Create Date: 2026-02-09
+
+This migration adds a `config` JSONB column to the banks table to support
+per-bank configuration overrides. This enables hierarchical configuration where:
+- Global config is loaded from environment variables
+- Tenant config is provided via TenantExtension
+- Bank config overrides are stored in banks.config JSONB column
+
+The config column stores overrides for hierarchical fields (LLM settings,
+retention parameters, retrieval settings, etc.) in Python field name format.
+"""
+
+from collections.abc import Sequence
+
+import sqlalchemy as sa
+from alembic import context, op
+from sqlalchemy.dialects.postgresql import JSONB
+
+revision: str = "x9s0t1u2v3w4"
+down_revision: str | Sequence[str] | None = "w8r9s0t1u2v3"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (required for multi-tenant support)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    """Add config JSONB column to banks table with GIN index."""
+    schema = _get_schema_prefix()
+
+    # Add config column to banks table
+    op.execute(f"""
+        ALTER TABLE {schema}banks
+        ADD COLUMN config JSONB NOT NULL DEFAULT '{{}}'::jsonb
+    """)
+
+    # Add GIN index for efficient JSONB queries
+    op.execute(f"""
+        CREATE INDEX idx_banks_config
+        ON {schema}banks
+        USING gin(config)
+    """)
+
+
+def downgrade() -> None:
+    """Remove config column and index from banks table."""
+    schema = _get_schema_prefix()
+
+    # Drop index first
+    op.execute(f"DROP INDEX IF EXISTS {schema}idx_banks_config")
+
+    # Drop column
+    op.execute(f"""
+        ALTER TABLE {schema}banks
+        DROP COLUMN IF EXISTS config
+    """)

hindsight-api/hindsight_api/alembic/versions/y0t1u2v3w4x5_add_result_metadata_gin_index.py

@@ -0,0 +1,49 @@
+"""Add GIN index on async_operations.result_metadata for parent_operation_id queries
+
+Revision ID: y0t1u2v3w4x5
+Revises: x9s0t1u2v3w4
+Create Date: 2026-02-13
+
+This migration adds a GIN index on the result_metadata JSONB column in the
+async_operations table to support efficient queries for child operations by
+parent_operation_id.
+
+The index enables fast lookups when querying for child operations:
+  SELECT * FROM async_operations
+  WHERE result_metadata::jsonb @> '{"parent_operation_id": "uuid"}'::jsonb
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+revision: str = "y0t1u2v3w4x5"
+down_revision: str | Sequence[str] | None = "x9s0t1u2v3w4"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (required for multi-tenant support)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    """Add GIN index on result_metadata for efficient parent_operation_id queries."""
+    schema = _get_schema_prefix()
+
+    # Add GIN index for JSONB containment queries (@> operator)
+    op.execute(f"""
+        CREATE INDEX idx_async_operations_result_metadata
+        ON {schema}async_operations
+        USING gin(result_metadata)
+    """)
+
+
+def downgrade() -> None:
+    """Remove GIN index on result_metadata."""
+    schema = _get_schema_prefix()
+
+    # Drop index
+    op.execute(f"DROP INDEX IF EXISTS {schema}idx_async_operations_result_metadata")

hindsight-api/hindsight_api/api/__init__.py

@@ -6,7 +6,6 @@ Provides both HTTP REST API and MCP (Model Context Protocol) server.
 
 import logging
 from contextlib import asynccontextmanager
-from typing import Optional
 
 from fastapi import FastAPI
 
@@ -46,14 +45,14 @@ def create_app(
         # Both HTTP and MCP
         app = create_app(memory, mcp_api_enabled=True)
     """
-    mcp_app = None
+    mcp_servers = None
 
-    # Create MCP app first if enabled (we need its lifespan for chaining)
+    # Create MCP servers first if enabled (we need their lifespans for chaining)
     if mcp_api_enabled:
         try:
-            from .mcp import create_mcp_app
+            from .mcp import MCPMiddleware, create_mcp_servers
 
-            mcp_app = create_mcp_app(memory=memory)
+            mcp_servers = create_mcp_servers(memory=memory)
         except ImportError as e:
             logger.error(f"MCP server requested but dependencies not available: {e}")
             logger.error("Install with: pip install hindsight-api[mcp]")
@@ -70,30 +69,41 @@ def create_app(
         app = FastAPI(title="Hindsight API", version="0.0.7")
         logger.info("HTTP REST API disabled")
 
-    # Mount MCP server and chain its lifespan if enabled
-    if mcp_app is not None:
-        # Get the MCP app's underlying Starlette app for lifespan access
-        mcp_starlette_app = mcp_app.mcp_app
+    # Add MCP middleware and chain its lifespan if enabled
+    if mcp_servers is not None:
+        multi_bank_server, single_bank_server, multi_bank_starlette_app, single_bank_starlette_app = mcp_servers
 
         # Store the original lifespan
         original_lifespan = app.router.lifespan_context
 
         @asynccontextmanager
         async def chained_lifespan(app_instance: FastAPI):
-            """Chain the MCP lifespan with the main app lifespan."""
-            # Start MCP lifespan first
-            async with mcp_starlette_app.router.lifespan_context(mcp_starlette_app):
-                logger.info("MCP lifespan started")
-                # Then start the original app lifespan
-                async with original_lifespan(app_instance):
-                    yield
-            logger.info("MCP lifespan stopped")
+            """Chain both MCP lifespans with the main app lifespan."""
+            # Start both MCP lifespans (multi-bank and single-bank)
+            async with multi_bank_starlette_app.router.lifespan_context(multi_bank_starlette_app):
+                async with single_bank_starlette_app.router.lifespan_context(single_bank_starlette_app):
+                    logger.info("MCP lifespans started (multi-bank and single-bank)")
+                    # Then start the original app lifespan
+                    async with original_lifespan(app_instance):
+                        yield
+                logger.info("MCP lifespans stopped")
 
         # Replace the app's lifespan with the chained version
         app.router.lifespan_context = chained_lifespan
 
-        # Mount the MCP middleware
-        app.mount(mcp_mount_path, mcp_app)
+        # Add MCP as a wrapping middleware — intercepts /mcp* requests directly,
+        # passes everything else through to the FastAPI app. No Starlette Mount
+        # means no 307 redirect for /mcp (no trailing slash).
+        app.add_middleware(
+            MCPMiddleware,
+            memory=memory,
+            prefix=mcp_mount_path,
+            multi_bank_app=multi_bank_starlette_app,
+            single_bank_app=single_bank_starlette_app,
+            multi_bank_server=multi_bank_server,
+            single_bank_server=single_bank_server,
+        )
+
         logger.info(f"MCP server enabled at {mcp_mount_path}/")
 
     return app

hindsight-api/hindsight_api/api/mcp.py

@@ -1,4 +1,4 @@
-"""Hindsight MCP Server implementation using FastMCP."""
+"""Hindsight MCP Server implementation using FastMCP (HTTP transport)."""
 
 import json
 import logging
@@ -8,7 +8,10 @@ from contextvars import ContextVar
 from fastmcp import FastMCP
 
 from hindsight_api import MemoryEngine
-from hindsight_api.engine.response_models import VALID_RECALL_FACT_TYPES
+from hindsight_api.engine.memory_engine import _current_schema
+from hindsight_api.extensions import MCPExtension, load_extension
+from hindsight_api.extensions.tenant import AuthenticationError
+from hindsight_api.mcp_tools import MCPToolsConfig, register_mcp_tools
 from hindsight_api.models import RequestContext
 
 # Configure logging from HINDSIGHT_API_LOG_LEVEL environment variable
@@ -30,21 +33,49 @@ logger = logging.getLogger(__name__)
 # Default bank_id from environment variable
 DEFAULT_BANK_ID = os.environ.get("HINDSIGHT_MCP_BANK_ID", "default")
 
+# Legacy MCP authentication token (for backwards compatibility)
+# If set, this token is checked first before TenantExtension auth
+MCP_AUTH_TOKEN = os.environ.get("HINDSIGHT_API_MCP_AUTH_TOKEN")
+
 # Context variable to hold the current bank_id
 _current_bank_id: ContextVar[str | None] = ContextVar("current_bank_id", default=None)
 
+# Context variable to hold the current API key (for tenant auth propagation)
+_current_api_key: ContextVar[str | None] = ContextVar("current_api_key", default=None)
+
+# Context variables for tenant_id and api_key_id (set by authenticate, used by usage metering)
+_current_tenant_id: ContextVar[str | None] = ContextVar("current_tenant_id", default=None)
+_current_api_key_id: ContextVar[str | None] = ContextVar("current_api_key_id", default=None)
+
 
 def get_current_bank_id() -> str | None:
     """Get the current bank_id from context."""
     return _current_bank_id.get()
 
 
-def create_mcp_server(memory: MemoryEngine) -> FastMCP:
+def get_current_api_key() -> str | None:
+    """Get the current API key from context."""
+    return _current_api_key.get()
+
+
+def get_current_tenant_id() -> str | None:
+    """Get the current tenant_id from context."""
+    return _current_tenant_id.get()
+
+
+def get_current_api_key_id() -> str | None:
+    """Get the current api_key_id from context."""
+    return _current_api_key_id.get()
+
+
+def create_mcp_server(memory: MemoryEngine, multi_bank: bool = True) -> FastMCP:
     """
     Create and configure the Hindsight MCP server.
 
     Args:
         memory: MemoryEngine instance (required)
+        multi_bank: If True, expose all tools with bank_id parameters (default).
+                   If False, only expose bank-scoped tools without bank_id parameters.
 
     Returns:
         Configured FastMCP server instance with stateless_http enabled
@@ -52,218 +83,137 @@ def create_mcp_server(memory: MemoryEngine) -> FastMCP:
     # Use stateless_http=True for Claude Code compatibility
     mcp = FastMCP("hindsight-mcp-server", stateless_http=True)
 
-    @mcp.tool()
-    async def retain(
-        content: str,
-        context: str = "general",
-        async_processing: bool = True,
-        bank_id: str | None = None,
-    ) -> str:
-        """
-        Store important information to long-term memory.
+    # Configure and register tools using shared module
+    config = MCPToolsConfig(
+        bank_id_resolver=get_current_bank_id,
+        api_key_resolver=get_current_api_key,  # Propagate API key for tenant auth
+        tenant_id_resolver=get_current_tenant_id,  # Propagate tenant_id for usage metering
+        api_key_id_resolver=get_current_api_key_id,  # Propagate api_key_id for usage metering
+        include_bank_id_param=multi_bank,
+        tools=None
+        if multi_bank
+        else {
+            "retain",
+            "recall",
+            "reflect",
+            "list_mental_models",
+            "get_mental_model",
+            "create_mental_model",
+            "update_mental_model",
+            "delete_mental_model",
+            "refresh_mental_model",
+        },  # Scoped tools for single-bank mode (excludes bank management: list_banks, create_bank)
+        retain_fire_and_forget=False,  # HTTP MCP supports sync/async modes
+    )
 
-        Use this tool PROACTIVELY whenever the user shares:
-        - Personal facts, preferences, or interests
-        - Important events or milestones
-        - User history, experiences, or background
-        - Decisions, opinions, or stated preferences
-        - Goals, plans, or future intentions
-        - Relationships or people mentioned
-        - Work context, projects, or responsibilities
+    register_mcp_tools(mcp, memory, config)
 
-        Args:
-            content: The fact/memory to store (be specific and include relevant details)
-            context: Category for the memory (e.g., 'preferences', 'work', 'hobbies', 'family'). Default: 'general'
-            async_processing: If True, queue for background processing and return immediately. If False, wait for completion. Default: True
-            bank_id: Optional bank to store in (defaults to session bank). Use for cross-bank operations.
-        """
-        try:
-            target_bank = bank_id or get_current_bank_id()
-            if target_bank is None:
-                return "Error: No bank_id configured"
-            contents = [{"content": content, "context": context}]
-            if async_processing:
-                # Queue for background processing and return immediately
-                result = await memory.submit_async_retain(
-                    bank_id=target_bank, contents=contents, request_context=RequestContext()
-                )
-                return f"Memory queued for background processing (operation_id: {result.get('operation_id', 'N/A')})"
-            else:
-                # Wait for completion
-                await memory.retain_batch_async(
-                    bank_id=target_bank,
-                    contents=contents,
-                    request_context=RequestContext(),
-                )
-                return f"Memory stored successfully in bank '{target_bank}'"
-        except Exception as e:
-            logger.error(f"Error storing memory: {e}", exc_info=True)
-            return f"Error: {str(e)}"
+    # Load and register additional tools from MCP extension if configured
+    mcp_extension = load_extension("MCP", MCPExtension)
+    if mcp_extension:
+        logger.info(f"Loading MCP extension: {mcp_extension.__class__.__name__}")
+        mcp_extension.register_tools(mcp, memory)
 
-    @mcp.tool()
-    async def recall(query: str, max_tokens: int = 4096, bank_id: str | None = None) -> str:
-        """
-        Search memories to provide personalized, context-aware responses.
-
-        Use this tool PROACTIVELY to:
-        - Check user's preferences before making suggestions
-        - Recall user's history to provide continuity
-        - Remember user's goals and context
-        - Personalize responses based on past interactions
-
-        Args:
-            query: Natural language search query (e.g., "user's food preferences", "what projects is user working on")
-            max_tokens: Maximum tokens in the response (default: 4096)
-            bank_id: Optional bank to search in (defaults to session bank). Use for cross-bank operations.
-        """
-        try:
-            target_bank = bank_id or get_current_bank_id()
-            if target_bank is None:
-                return "Error: No bank_id configured"
-            from hindsight_api.engine.memory_engine import Budget
-
-            recall_result = await memory.recall_async(
-                bank_id=target_bank,
-                query=query,
-                fact_type=list(VALID_RECALL_FACT_TYPES),
-                budget=Budget.HIGH,
-                max_tokens=max_tokens,
-                request_context=RequestContext(),
-            )
-
-            # Use model's JSON serialization
-            return recall_result.model_dump_json(indent=2)
-        except Exception as e:
-            logger.error(f"Error searching: {e}", exc_info=True)
-            return f'{{"error": "{e}", "results": []}}'
-
-    @mcp.tool()
-    async def reflect(query: str, context: str | None = None, budget: str = "low", bank_id: str | None = None) -> str:
-        """
-        Generate thoughtful analysis by synthesizing stored memories with the bank's personality.
-
-        WHEN TO USE THIS TOOL:
-        Use reflect when you need reasoned analysis, not just fact retrieval. This tool
-        thinks through the question using everything the bank knows and its personality traits.
-
-        EXAMPLES OF GOOD QUERIES:
-        - "What patterns have emerged in how I approach debugging?"
-        - "Based on my past decisions, what architectural style do I prefer?"
-        - "What might be the best approach for this problem given what you know about me?"
-        - "How should I prioritize these tasks based on my goals?"
-
-        HOW IT DIFFERS FROM RECALL:
-        - recall: Returns raw facts matching your search (fast lookup)
-        - reflect: Reasons across memories to form a synthesized answer (deeper analysis)
-
-        Use recall for "what did I say about X?" and reflect for "what should I do about X?"
-
-        Args:
-            query: The question or topic to reflect on
-            context: Optional context about why this reflection is needed
-            budget: Search budget - 'low', 'mid', or 'high' (default: 'low')
-            bank_id: Optional bank to reflect in (defaults to session bank). Use for cross-bank operations.
-        """
-        try:
-            target_bank = bank_id or get_current_bank_id()
-            if target_bank is None:
-                return "Error: No bank_id configured"
-            from hindsight_api.engine.memory_engine import Budget
-
-            # Map string budget to enum
-            budget_map = {"low": Budget.LOW, "mid": Budget.MID, "high": Budget.HIGH}
-            budget_enum = budget_map.get(budget.lower(), Budget.LOW)
-
-            reflect_result = await memory.reflect_async(
-                bank_id=target_bank,
-                query=query,
-                budget=budget_enum,
-                context=context,
-                request_context=RequestContext(),
-            )
-
-            return reflect_result.model_dump_json(indent=2)
-        except Exception as e:
-            logger.error(f"Error reflecting: {e}", exc_info=True)
-            return f'{{"error": "{e}", "text": ""}}'
-
-    @mcp.tool()
-    async def list_banks() -> str:
-        """
-        List all available memory banks.
-
-        Use this tool to discover what memory banks exist in the system.
-        Each bank is an isolated memory store (like a separate "brain").
-
-        Returns:
-            JSON list of banks with their IDs, names, dispositions, and backgrounds.
-        """
-        try:
-            banks = await memory.list_banks(request_context=RequestContext())
-            return json.dumps({"banks": banks}, indent=2)
-        except Exception as e:
-            logger.error(f"Error listing banks: {e}", exc_info=True)
-            return f'{{"error": "{e}", "banks": []}}'
-
-    @mcp.tool()
-    async def create_bank(bank_id: str, name: str | None = None, background: str | None = None) -> str:
-        """
-        Create a new memory bank or get an existing one.
-
-        Memory banks are isolated stores - each one is like a separate "brain" for a user/agent.
-        Banks are auto-created with default settings if they don't exist.
-
-        Args:
-            bank_id: Unique identifier for the bank (e.g., 'user-123', 'agent-alpha')
-            name: Optional human-friendly name for the bank
-            background: Optional background context about the bank's owner/purpose
-        """
-        try:
-            # get_bank_profile auto-creates bank if it doesn't exist
-            profile = await memory.get_bank_profile(bank_id, request_context=RequestContext())
-
-            # Update name/background if provided
-            if name is not None or background is not None:
-                await memory.update_bank(
-                    bank_id,
-                    name=name,
-                    background=background,
-                    request_context=RequestContext(),
-                )
-                # Fetch updated profile
-                profile = await memory.get_bank_profile(bank_id, request_context=RequestContext())
-
-            # Serialize disposition if it's a Pydantic model
-            if "disposition" in profile and hasattr(profile["disposition"], "model_dump"):
-                profile["disposition"] = profile["disposition"].model_dump()
-            return json.dumps(profile, indent=2)
-        except Exception as e:
-            logger.error(f"Error creating bank: {e}", exc_info=True)
-            return f'{{"error": "{e}"}}'
+    # Make all tools tolerant of extra arguments from LLMs (e.g., "explanation")
+    _make_tools_tolerant(mcp)
 
     return mcp
 
 
+def _make_tools_tolerant(mcp: FastMCP) -> None:
+    """Wrap all tool run methods to strip unknown arguments before validation.
+
+    LLMs frequently add extra fields like "explanation" or "reasoning" to tool calls.
+    FastMCP's Pydantic TypeAdapter rejects these with "Unexpected keyword argument".
+    This wraps each tool's run() to filter arguments to only known parameters.
+    """
+    try:
+        for name, tool in mcp._tool_manager._tools.items():
+            if hasattr(tool, "parameters") and tool.parameters:
+                allowed = set(tool.parameters.get("properties", {}).keys())
+                original_run = tool.run
+
+                async def _tolerant_run(arguments, _allowed=allowed, _orig=original_run):
+                    extra_keys = set(arguments.keys()) - _allowed
+                    if extra_keys:
+                        logger.debug(f"Stripping unknown arguments from tool call: {extra_keys}")
+                        arguments = {k: v for k, v in arguments.items() if k in _allowed}
+                    return await _orig(arguments)
+
+                # FunctionTool is a Pydantic model with extra='forbid', so use
+                # object.__setattr__ to bypass Pydantic's setter validation.
+                object.__setattr__(tool, "run", _tolerant_run)
+    except (AttributeError, KeyError) as e:
+        logger.warning(f"Could not make tools tolerant of extra arguments: {e}")
+
+
 class MCPMiddleware:
-    """ASGI middleware that extracts bank_id from header or path and sets context.
+    """ASGI middleware that intercepts MCP requests and routes to appropriate MCP server.
 
-    Bank ID can be provided via:
-    1. X-Bank-Id header (recommended for Claude Code)
-    2. URL path: /mcp/{bank_id}/
-    3. Environment variable HINDSIGHT_MCP_BANK_ID (fallback default)
+    This middleware wraps the main FastAPI app and intercepts requests matching the
+    configured prefix (default: /mcp). Non-MCP requests pass through to the inner app.
 
-    For Claude Code, configure with:
+    Authentication:
+        1. If HINDSIGHT_API_MCP_AUTH_TOKEN is set (legacy), validates against that token
+        2. Otherwise, uses TenantExtension.authenticate_mcp() from the MemoryEngine
+           - DefaultTenantExtension: no auth required (local dev)
+           - ApiKeyTenantExtension: validates against env var
+
+    Two modes based on URL structure:
+
+    1. Multi-bank mode (for /mcp/ root endpoint):
+       - Exposes all tools: retain, recall, reflect, list_banks, create_bank
+       - All tools include optional bank_id parameter for cross-bank operations
+       - Bank ID from: X-Bank-Id header or HINDSIGHT_MCP_BANK_ID env var
+
+    2. Single-bank mode (for /mcp/{bank_id}/ endpoints):
+       - Exposes bank-scoped tools only: retain, recall, reflect
+       - No bank_id parameter (comes from URL)
+       - No bank management tools (list_banks, create_bank)
+       - Recommended for agent isolation
+
+    Bank ID resolution priority:
+        1. URL path (e.g., /mcp/{bank_id}/) → single-bank mode
+        2. X-Bank-Id header → multi-bank mode
+        3. HINDSIGHT_MCP_BANK_ID env var → multi-bank mode (default: "default")
+
+    Examples:
+        # Single-bank mode (recommended for agent isolation)
+        claude mcp add --transport http my-agent http://localhost:8888/mcp/my-agent-bank/ \\
+            --header "Authorization: Bearer <token>"
+
+        # Multi-bank mode (for cross-bank operations)
         claude mcp add --transport http hindsight http://localhost:8888/mcp \\
-            --header "X-Bank-Id: my-bank"
+            --header "X-Bank-Id: my-bank" --header "Authorization: Bearer <token>"
     """
 
-    def __init__(self, app, memory: MemoryEngine):
+    def __init__(
+        self,
+        app,
+        memory: MemoryEngine,
+        prefix: str = "/mcp",
+        multi_bank_app=None,
+        single_bank_app=None,
+        multi_bank_server=None,
+        single_bank_server=None,
+    ):
         self.app = app
+        self.prefix = prefix
         self.memory = memory
-        self.mcp_server = create_mcp_server(memory)
-        self.mcp_app = self.mcp_server.http_app(path="/")
-        # Expose the lifespan for the parent app to chain
-        self.lifespan = self.mcp_app.lifespan_handler if hasattr(self.mcp_app, "lifespan_handler") else None
+        self.tenant_extension = memory._tenant_extension
+
+        if multi_bank_app and single_bank_app:
+            # Pre-created servers (used when called via add_middleware from create_app)
+            self.multi_bank_app = multi_bank_app
+            self.single_bank_app = single_bank_app
+            self.multi_bank_server = multi_bank_server
+            self.single_bank_server = single_bank_server
+        else:
+            # Create servers internally (for direct construction / tests)
+            self.multi_bank_server = create_mcp_server(memory, multi_bank=True)
+            self.multi_bank_app = self.multi_bank_server.http_app(path="/")
+            self.single_bank_server = create_mcp_server(memory, multi_bank=False)
+            self.single_bank_app = self.single_bank_server.http_app(path="/")
 
     def _get_header(self, scope: dict, name: str) -> str | None:
         """Extract a header value from ASGI scope."""
@@ -275,54 +225,113 @@ class MCPMiddleware:
 
     async def __call__(self, scope, receive, send):
         if scope["type"] != "http":
-            await self.mcp_app(scope, receive, send)
+            await self.app(scope, receive, send)
             return
 
         path = scope.get("path", "")
 
-        # Strip any mount prefix (e.g., /mcp) that FastAPI might not have stripped
-        root_path = scope.get("root_path", "")
-        if root_path and path.startswith(root_path):
-            path = path[len(root_path) :] or "/"
+        # Check if this is an MCP request (matches prefix)
+        if not (path == self.prefix or path.startswith(self.prefix + "/")):
+            # Not an MCP request — pass through to the inner app
+            await self.app(scope, receive, send)
+            return
 
-        # Also handle case where mount path wasn't stripped (e.g., /mcp/...)
-        if path.startswith("/mcp/"):
-            path = path[4:]  # Remove /mcp prefix
-        elif path == "/mcp":
-            path = "/"
+        # Strip prefix from path
+        path = path[len(self.prefix) :] or "/"
 
-        # Try to get bank_id from header first (for Claude Code compatibility)
-        bank_id = self._get_header(scope, "X-Bank-Id")
+        # Extract auth token from header (for tenant auth propagation)
+        auth_header = self._get_header(scope, "Authorization")
+        auth_token: str | None = None
+        if auth_header:
+            # Support both "Bearer <token>" and direct token
+            auth_token = auth_header[7:].strip() if auth_header.startswith("Bearer ") else auth_header.strip()
 
-        # MCP endpoint paths that should not be treated as bank_ids
-        MCP_ENDPOINTS = {"sse", "messages"}
+        # Authenticate: check legacy MCP_AUTH_TOKEN first, then TenantExtension
+        tenant_context = None
+        auth_tenant_id: str | None = None
+        auth_api_key_id: str | None = None
+        if MCP_AUTH_TOKEN:
+            # Legacy authentication mode - validate against static token
+            if not auth_token:
+                await self._send_error(send, 401, "Authorization header required")
+                return
+            if auth_token != MCP_AUTH_TOKEN:
+                await self._send_error(send, 401, "Invalid authentication token")
+                return
+            # Legacy mode doesn't use tenant schemas
+            tenant_context = None
+        else:
+            # Use TenantExtension.authenticate_mcp() for auth
+            try:
+                auth_context = RequestContext(api_key=auth_token)
+                tenant_context = await self.tenant_extension.authenticate_mcp(auth_context)
+                # Capture tenant_id and api_key_id set by authenticate() for usage metering
+                auth_tenant_id = auth_context.tenant_id
+                auth_api_key_id = auth_context.api_key_id
+            except AuthenticationError as e:
+                await self._send_error(send, 401, str(e))
+                return
 
-        # If no header, try to extract from path: /{bank_id}/...
+        # Set schema from tenant context so downstream DB queries use the correct schema
+        schema_token = (
+            _current_schema.set(tenant_context.schema_name) if tenant_context and tenant_context.schema_name else None
+        )
+
+        # Resolve bank_id: path takes priority over header.
+        # Path = user's explicit connection endpoint (e.g., /mcp/my-bank/).
+        # X-Bank-Id header = per-request override for multi-bank mode only.
+        bank_id = None
+        bank_id_from_path = False
         new_path = path
-        if not bank_id and path.startswith("/") and len(path) > 1:
+
+        # First, try to extract from path: /{bank_id}/...
+        if path.startswith("/") and len(path) > 1:
             parts = path[1:].split("/", 1)
-            # Don't treat MCP endpoints as bank_ids
-            if parts[0] and parts[0] not in MCP_ENDPOINTS:
-                # First segment looks like a bank_id
+            if parts[0]:
                 bank_id = parts[0]
+                bank_id_from_path = True
                 new_path = "/" + parts[1] if len(parts) > 1 else "/"
 
+        # If no path-based bank_id, try X-Bank-Id header (multi-bank mode)
+        if not bank_id:
+            bank_id = self._get_header(scope, "X-Bank-Id")
+
         # Fall back to default bank_id
         if not bank_id:
             bank_id = DEFAULT_BANK_ID
             logger.debug(f"Using default bank_id: {bank_id}")
 
-        # Set bank_id context
-        token = _current_bank_id.set(bank_id)
+        # Select the appropriate MCP app based on how bank_id was provided:
+        # - Path-based bank_id → single-bank app (no bank_id param, scoped tools)
+        # - Header/env bank_id → multi-bank app (bank_id param, all tools)
+        target_app = self.single_bank_app if bank_id_from_path else self.multi_bank_app
+
+        # Set bank_id, api_key, tenant_id, and api_key_id context
+        bank_id_token = _current_bank_id.set(bank_id)
+        # Store the auth token for tenant extension to validate
+        api_key_token = _current_api_key.set(auth_token) if auth_token else None
+        # Store tenant_id and api_key_id from authentication for usage metering
+        tenant_id_token = _current_tenant_id.set(auth_tenant_id) if auth_tenant_id else None
+        api_key_id_token = _current_api_key_id.set(auth_api_key_id) if auth_api_key_id else None
         try:
             new_scope = scope.copy()
             new_scope["path"] = new_path
             # Clear root_path since we're passing directly to the app
             new_scope["root_path"] = ""
 
-            # Wrap send to rewrite the SSE endpoint URL to include bank_id if using path-based routing
+            # Wrap send to rewrite the SSE endpoint URL to include bank_id if using path-based routing.
+            # Only rewrite SSE (text/event-stream) responses to avoid corrupting tool results
+            # that might contain the literal string "data: /messages".
+            is_sse_response = False
+
             async def send_wrapper(message):
-                if message["type"] == "http.response.body":
+                nonlocal is_sse_response
+                if message["type"] == "http.response.start":
+                    for header_name, header_value in message.get("headers", []):
+                        if header_name == b"content-type" and b"text/event-stream" in header_value:
+                            is_sse_response = True
+                            break
+                if message["type"] == "http.response.body" and bank_id_from_path and is_sse_response:
                     body = message.get("body", b"")
                     if body and b"/messages" in body:
                         # Rewrite /messages to /{bank_id}/messages in SSE endpoint event
@@ -330,9 +339,17 @@ class MCPMiddleware:
                         message = {**message, "body": body}
                 await send(message)
 
-            await self.mcp_app(new_scope, receive, send_wrapper)
+            await target_app(new_scope, receive, send_wrapper)
         finally:
-            _current_bank_id.reset(token)
+            _current_bank_id.reset(bank_id_token)
+            if api_key_token is not None:
+                _current_api_key.reset(api_key_token)
+            if tenant_id_token is not None:
+                _current_tenant_id.reset(tenant_id_token)
+            if api_key_id_token is not None:
+                _current_api_key_id.reset(api_key_id_token)
+            if schema_token is not None:
+                _current_schema.reset(schema_token)
 
     async def _send_error(self, send, status: int, message: str):
         """Send an error response."""
@@ -352,19 +369,19 @@ class MCPMiddleware:
         )
 
 
-def create_mcp_app(memory: MemoryEngine):
-    """
-    Create an ASGI app that handles MCP requests.
+def create_mcp_servers(memory: MemoryEngine):
+    """Create multi-bank and single-bank MCP servers and their Starlette apps.
 
-    Bank ID can be provided via:
-    1. X-Bank-Id header: claude mcp add --transport http hindsight http://localhost:8888/mcp --header "X-Bank-Id: my-bank"
-    2. URL path: /mcp/{bank_id}/
-    3. Environment variable HINDSIGHT_MCP_BANK_ID (fallback, default: "default")
-
-    Args:
-        memory: MemoryEngine instance
+    Returns the servers and apps separately so lifespans can be chained before
+    the middleware wraps the main app.
 
     Returns:
-        ASGI application
+        Tuple of (multi_bank_server, single_bank_server, multi_bank_app, single_bank_app)
     """
-    return MCPMiddleware(None, memory)
+    multi_bank_server = create_mcp_server(memory, multi_bank=True)
+    multi_bank_app = multi_bank_server.http_app(path="/")
+
+    single_bank_server = create_mcp_server(memory, multi_bank=False)
+    single_bank_app = single_bank_server.http_app(path="/")
+
+    return multi_bank_server, single_bank_server, multi_bank_app, single_bank_app

hindsight-api/hindsight_api/banner.py

@@ -4,6 +4,8 @@ Banner display for Hindsight API startup.
 Shows the logo and tagline with gradient colors.
 """
 
+from .utils import mask_network_location
+
 # Gradient colors: #0074d9 -> #009296
 GRADIENT_START = (0, 116, 217)  # #0074d9
 GRADIENT_END = (0, 146, 150)  # #009296
@@ -83,14 +85,21 @@ def print_startup_info(
     embeddings_provider: str,
     reranker_provider: str,
     mcp_enabled: bool = False,
+    version: str | None = None,
+    vector_extension: str | None = None,
+    text_search_extension: str | None = None,
 ):
     """Print styled startup information."""
     print(color_start("Starting Hindsight API..."))
+    if version:
+        print(f"  {dim('Version:')} {color(f'v{version}', 0.1)}")
     print(f"  {dim('URL:')} {color(f'http://{host}:{port}', 0.2)}")
-    print(f"  {dim('Database:')} {color(database_url, 0.4)}")
+    print(f"  {dim('Database:')} {color(mask_network_location(database_url), 0.4)}")
     print(f"  {dim('LLM:')} {color(f'{llm_provider} / {llm_model}', 0.6)}")
     print(f"  {dim('Embeddings:')} {color(embeddings_provider, 0.8)}")
     print(f"  {dim('Reranker:')} {color(reranker_provider, 1.0)}")
+    extensions = f"{vector_extension or 'default'} (vector) / {text_search_extension or 'default'} (text)"
+    print(f"  {dim('Extensions:')} {color(extensions, 0.4)}")
     if mcp_enabled:
         print(f"  {dim('MCP:')} {color_end('enabled at /mcp')}")
     print()

hindsight-api/hindsight_api/config_resolver.py

@@ -0,0 +1,274 @@
+"""
+Configuration resolution with hierarchical overrides.
+
+Resolves config values through the hierarchy:
+  Global (env vars) → Tenant config (via extension) → Bank config (database)
+
+Config values are resolved on every request to ensure consistency across
+multiple API servers.
+"""
+
+import json
+import logging
+from dataclasses import asdict
+from typing import Any
+
+import asyncpg
+
+from hindsight_api.config import HindsightConfig, _get_raw_config, normalize_config_dict
+from hindsight_api.extensions.tenant import TenantExtension
+from hindsight_api.models import RequestContext
+
+logger = logging.getLogger(__name__)
+
+
+class ConfigResolver:
+    """Resolves hierarchical configuration with tenant/bank overrides."""
+
+    def __init__(self, pool: asyncpg.Pool, tenant_extension: TenantExtension | None = None):
+        """
+        Initialize config resolver.
+
+        Args:
+            pool: Database connection pool
+            tenant_extension: Optional tenant extension for tenant-level config and permissions
+        """
+        self.pool = pool
+        self.tenant_extension = tenant_extension
+        self._global_config = _get_raw_config()
+        self._configurable_fields = HindsightConfig.get_configurable_fields()
+        self._credential_fields = HindsightConfig.get_credential_fields()
+
+    async def resolve_full_config(self, bank_id: str, context: RequestContext | None = None) -> HindsightConfig:
+        """
+        Resolve full HindsightConfig for a bank with hierarchical overrides applied.
+
+        This is for INTERNAL USE ONLY. Returns the complete config object with all fields
+        including credentials and static fields. Use get_bank_config() for API responses.
+
+        Resolution order:
+        1. Global config (from environment variables)
+        2. Tenant config overrides (from TenantExtension.get_tenant_config())
+        3. Bank config overrides (from banks.config JSONB)
+
+        Args:
+            bank_id: Bank identifier
+            context: Request context for tenant config resolution
+
+        Returns:
+            Complete HindsightConfig with hierarchical overrides applied
+        """
+        # Start with global config (all fields)
+        config_dict = asdict(self._global_config)
+
+        # Load tenant config overrides (if tenant extension available)
+        if self.tenant_extension and context:
+            try:
+                tenant_overrides = await self.tenant_extension.get_tenant_config(context)
+                if tenant_overrides:
+                    # Normalize keys and filter to configurable fields only
+                    normalized_tenant = normalize_config_dict(tenant_overrides)
+                    configurable_tenant = {k: v for k, v in normalized_tenant.items() if k in self._configurable_fields}
+                    config_dict.update(configurable_tenant)
+                    logger.debug(
+                        f"Applied tenant config overrides for bank {bank_id}: {list(configurable_tenant.keys())}"
+                    )
+            except Exception as e:
+                logger.warning(f"Failed to load tenant config for bank {bank_id}: {e}")
+
+        # Load bank config overrides
+        bank_overrides = await self._load_bank_config(bank_id)
+        if bank_overrides:
+            config_dict.update(bank_overrides)
+            logger.debug(f"Applied bank config overrides for bank {bank_id}: {list(bank_overrides.keys())}")
+
+        # Return full config object (dataclass doesn't have __init__ that accepts kwargs, so we update the object)
+        # Create a new config instance by copying the global config and updating fields
+        resolved_config = HindsightConfig(**config_dict)
+        return resolved_config
+
+    async def get_bank_config(self, bank_id: str, context: RequestContext | None = None) -> dict[str, Any]:
+        """
+        Get fully resolved config for a bank (filtered by permissions).
+
+        Resolution order:
+        1. Global config (from environment variables)
+        2. Tenant config overrides (from TenantExtension.get_tenant_config())
+        3. Bank config overrides (from banks.config JSONB)
+
+        Note: Config is resolved on every call (not cached) to ensure consistency
+        across multiple API servers.
+
+        SECURITY:
+        - Only returns configurable fields (excludes static/infrastructure fields)
+        - Filters out ALL credential fields (API keys, base URLs, etc.)
+        - Further filtered by tenant/bank permissions if extension provides them
+
+        Args:
+            bank_id: Bank identifier
+            context: Request context for tenant config resolution and permissions
+
+        Returns:
+            Dict of allowed configurable fields only (never includes credentials or static fields)
+        """
+        # Resolve full config with all hierarchical overrides
+        resolved_config = await self.resolve_full_config(bank_id, context)
+        config_dict = asdict(resolved_config)
+
+        # SECURITY: Filter to only configurable fields (exclude static/infrastructure)
+        filtered = {k: v for k, v in config_dict.items() if k in self._configurable_fields}
+
+        # SECURITY: Remove ALL credential fields (API keys, base URLs, etc.)
+        filtered = {k: v for k, v in filtered.items() if k not in self._credential_fields}
+
+        # PERMISSIONS: Further filter based on tenant/bank permissions
+        if self.tenant_extension and context:
+            try:
+                allowed_fields = await self.tenant_extension.get_allowed_config_fields(context, bank_id)
+                if allowed_fields is not None:  # None means "allow all"
+                    filtered = {k: v for k, v in filtered.items() if k in allowed_fields}
+                    logger.debug(
+                        f"Applied permission filter for bank {bank_id}: allowed={len(allowed_fields)} fields, "
+                        f"returned={len(filtered)} fields"
+                    )
+            except Exception as e:
+                logger.warning(f"Failed to load permissions for bank {bank_id}: {e}")
+
+        return filtered
+
+    async def _load_bank_config(self, bank_id: str) -> dict[str, Any]:
+        """
+        Load bank config overrides from banks.config JSONB column.
+
+        Args:
+            bank_id: Bank identifier
+
+        Returns:
+            Dict of config overrides (only configurable fields, normalized keys)
+        """
+        try:
+            async with self.pool.acquire() as conn:
+                row = await conn.fetchrow(
+                    """
+                    SELECT config FROM banks WHERE bank_id = $1
+                    """,
+                    bank_id,
+                )
+
+                if row and row["config"]:
+                    config_data = row["config"]
+
+                    # Handle case where JSONB is returned as JSON string
+                    if isinstance(config_data, str):
+                        config_data = json.loads(config_data)
+
+                    # Normalize keys (handle both env var format and Python field format)
+                    normalized = normalize_config_dict(config_data)
+
+                    # Only return overrides for configurable fields
+                    return {k: v for k, v in normalized.items() if k in self._configurable_fields}
+        except Exception as e:
+            logger.error(f"Failed to load bank config for {bank_id}: {e}")
+
+        return {}
+
+    async def update_bank_config(
+        self, bank_id: str, updates: dict[str, Any], context: RequestContext | None = None
+    ) -> None:
+        """
+        Update bank configuration overrides (with permission checking).
+
+        Args:
+            bank_id: Bank identifier
+            updates: Dict of config field names to new values.
+                    Keys can be in env var format (HINDSIGHT_API_LLM_PROVIDER)
+                    or Python field format (llm_provider).
+                    Only configurable fields are allowed.
+            context: Request context for permission checking
+
+        Raises:
+            ValueError: If attempting to override invalid/disallowed fields
+        """
+        # Normalize keys
+        normalized_updates = normalize_config_dict(updates)
+
+        # SECURITY: Reject credential fields explicitly
+        credential_attempts = set(normalized_updates.keys()) & self._credential_fields
+        if credential_attempts:
+            raise ValueError(
+                f"Cannot set credential fields via API: {sorted(credential_attempts)}. "
+                f"Credentials (API keys, base URLs) must be set at server level only."
+            )
+
+        # Validate all fields are configurable
+        invalid_fields = set(normalized_updates.keys()) - self._configurable_fields
+        if invalid_fields:
+            static_fields = HindsightConfig.get_static_fields()
+            invalid_static = invalid_fields & static_fields
+            if invalid_static:
+                raise ValueError(
+                    f"Cannot override static (server-level) fields: {sorted(invalid_static)}. "
+                    f"Only configurable fields can be overridden per-bank. "
+                    f"Configurable fields include: {sorted(list(self._configurable_fields)[:10])}... "
+                    f"(total: {len(self._configurable_fields)} fields)"
+                )
+            else:
+                raise ValueError(
+                    f"Unknown configuration fields: {sorted(invalid_fields)}. "
+                    f"Valid configurable fields: {sorted(list(self._configurable_fields)[:10])}..."
+                )
+
+        # PERMISSIONS: Check tenant/bank permissions
+        if self.tenant_extension and context:
+            try:
+                allowed_fields = await self.tenant_extension.get_allowed_config_fields(context, bank_id)
+                if allowed_fields is not None:  # None means "allow all"
+                    disallowed = set(normalized_updates.keys()) - allowed_fields
+                    if disallowed:
+                        raise ValueError(
+                            f"Not allowed to modify fields: {sorted(disallowed)}. "
+                            f"Your permissions allow: {sorted(list(allowed_fields)[:10])}..."
+                            if allowed_fields
+                            else "Not allowed to modify fields: {sorted(disallowed)}. "
+                            "Your permissions do not allow any config modifications."
+                        )
+            except ValueError:
+                raise  # Re-raise permission errors
+            except Exception as e:
+                logger.warning(f"Failed to check permissions for bank {bank_id}: {e}")
+                # Continue without permission check (fail open for backward compatibility)
+
+        # Merge with existing config (JSONB || operator)
+        async with self.pool.acquire() as conn:
+            await conn.execute(
+                """
+                UPDATE banks
+                SET config = config || $1::jsonb,
+                    updated_at = now()
+                WHERE bank_id = $2
+                """,
+                json.dumps(normalized_updates),
+                bank_id,
+            )
+
+        logger.info(f"Updated bank config for {bank_id}: {list(normalized_updates.keys())}")
+
+    async def reset_bank_config(self, bank_id: str) -> None:
+        """
+        Reset bank configuration to defaults (remove all overrides).
+
+        Args:
+            bank_id: Bank identifier
+        """
+        async with self.pool.acquire() as conn:
+            await conn.execute(
+                """
+                UPDATE banks
+                SET config = '{}'::jsonb,
+                    updated_at = now()
+                WHERE bank_id = $1
+                """,
+                bank_id,
+            )
+
+        logger.info(f"Reset bank config for {bank_id} to defaults")

hindsight-api/hindsight_api/daemon.py

@@ -1,11 +1,10 @@
 """
 Daemon mode support for Hindsight API.
 
-Provides idle timeout and lockfile management for running as a background daemon.
+Provides idle timeout for running as a background daemon.
 """
 
 import asyncio
-import fcntl
 import logging
 import os
 import sys
@@ -15,10 +14,11 @@ from pathlib import Path
 logger = logging.getLogger(__name__)
 
 # Default daemon configuration
-DEFAULT_DAEMON_PORT = 8889
+DEFAULT_DAEMON_PORT = 8888
 DEFAULT_IDLE_TIMEOUT = 0  # 0 = no auto-exit (hindsight-embed passes its own timeout)
-LOCKFILE_PATH = Path.home() / ".hindsight" / "daemon.lock"
-DAEMON_LOG_PATH = Path.home() / ".hindsight" / "daemon.log"
+
+# Allow override via environment variable for profile-specific logs
+DAEMON_LOG_PATH = Path(os.getenv("HINDSIGHT_API_DAEMON_LOG", str(Path.home() / ".hindsight" / "daemon.log")))
 
 
 class IdleTimeoutMiddleware:
@@ -52,82 +52,10 @@ class IdleTimeoutMiddleware:
                 logger.info(f"Idle timeout reached ({self.idle_timeout}s), shutting down daemon")
                 # Give a moment for any in-flight requests
                 await asyncio.sleep(1)
-                os._exit(0)
+                # Send SIGTERM to ourselves to trigger graceful shutdown
+                import signal
 
-
-class DaemonLock:
-    """
-    File-based lock to prevent multiple daemon instances.
-
-    Uses fcntl.flock for atomic locking on Unix systems.
-    """
-
-    def __init__(self, lockfile: Path = LOCKFILE_PATH):
-        self.lockfile = lockfile
-        self._fd = None
-
-    def acquire(self) -> bool:
-        """
-        Try to acquire the daemon lock.
-
-        Returns True if lock acquired, False if another daemon is running.
-        """
-        self.lockfile.parent.mkdir(parents=True, exist_ok=True)
-
-        try:
-            self._fd = open(self.lockfile, "w")
-            fcntl.flock(self._fd.fileno(), fcntl.LOCK_EX | fcntl.LOCK_NB)
-            # Write PID for debugging
-            self._fd.write(str(os.getpid()))
-            self._fd.flush()
-            return True
-        except (IOError, OSError):
-            # Lock is held by another process
-            if self._fd:
-                self._fd.close()
-                self._fd = None
-            return False
-
-    def release(self):
-        """Release the daemon lock."""
-        if self._fd:
-            try:
-                fcntl.flock(self._fd.fileno(), fcntl.LOCK_UN)
-                self._fd.close()
-            except Exception:
-                pass
-            finally:
-                self._fd = None
-                # Remove lockfile
-                try:
-                    self.lockfile.unlink()
-                except Exception:
-                    pass
-
-    def is_locked(self) -> bool:
-        """Check if the lock is held by another process."""
-        if not self.lockfile.exists():
-            return False
-
-        try:
-            fd = open(self.lockfile, "r")
-            fcntl.flock(fd.fileno(), fcntl.LOCK_EX | fcntl.LOCK_NB)
-            # We got the lock, so no one else has it
-            fcntl.flock(fd.fileno(), fcntl.LOCK_UN)
-            fd.close()
-            return False
-        except (IOError, OSError):
-            return True
-
-    def get_pid(self) -> int | None:
-        """Get the PID of the daemon holding the lock."""
-        if not self.lockfile.exists():
-            return None
-        try:
-            with open(self.lockfile, "r") as f:
-                return int(f.read().strip())
-        except (ValueError, IOError):
-            return None
+                os.kill(os.getpid(), signal.SIGTERM)
 
 
 def daemonize():
@@ -136,16 +64,21 @@ def daemonize():
 
     Uses double-fork technique to properly detach from terminal.
     """
-    # First fork
-    pid = os.fork()
-    if pid > 0:
-        # Parent exits
-        sys.exit(0)
+    # First fork - detach from parent
+    try:
+        pid = os.fork()
+        if pid > 0:
+            sys.exit(0)
+    except OSError as e:
+        sys.stderr.write(f"fork #1 failed: {e}\n")
+        sys.exit(1)
 
-    # Create new session
+    # Decouple from parent environment
+    os.chdir("/")
     os.setsid()
+    os.umask(0)
 
-    # Second fork to prevent zombie processes
+    # Second fork - prevent zombie
     pid = os.fork()
     if pid > 0:
         sys.exit(0)
@@ -178,27 +111,3 @@ def check_daemon_running(port: int = DEFAULT_DAEMON_PORT) -> bool:
         return result == 0
     except Exception:
         return False
-
-
-def stop_daemon(port: int = DEFAULT_DAEMON_PORT) -> bool:
-    """Stop a running daemon by sending SIGTERM to the process."""
-    lock = DaemonLock()
-    pid = lock.get_pid()
-
-    if pid is None:
-        return False
-
-    try:
-        import signal
-
-        os.kill(pid, signal.SIGTERM)
-        # Wait for process to exit
-        for _ in range(50):  # Wait up to 5 seconds
-            time.sleep(0.1)
-            try:
-                os.kill(pid, 0)  # Check if process exists
-            except OSError:
-                return True  # Process exited
-        return False
-    except OSError:
-        return False

hindsight-api/hindsight_api/engine/consolidation/__init__.py

@@ -0,0 +1,5 @@
+"""Consolidation engine for automatic learning creation from memories."""
+
+from .consolidator import run_consolidation_job
+
+__all__ = ["run_consolidation_job"]

hindsight-api/hindsight_api/engine/consolidation/prompts.py

@@ -0,0 +1,85 @@
+"""Prompts for the consolidation engine."""
+
+CONSOLIDATION_SYSTEM_PROMPT = """You are a memory consolidation system. Your job is to convert facts into durable knowledge (observations) and merge with existing knowledge when appropriate.
+
+You must output ONLY valid JSON with no markdown code blocks or additional text. However, the "text" field within each observation should use markdown formatting (headers, lists, bold, etc.) for clarity and readability.
+
+## EXTRACT DURABLE KNOWLEDGE, NOT EPHEMERAL STATE
+Facts often describe events or actions. Extract the DURABLE KNOWLEDGE implied by the fact, not the transient state.
+
+Examples of extracting durable knowledge:
+- "User moved to Room 203" -> "Room 203 exists" (location exists, not where user is now)
+- "User visited Acme Corp at Room 105" -> "Acme Corp is located in Room 105"
+- "User took the elevator to floor 3" -> "Floor 3 is accessible by elevator"
+- "User met Sarah at the lobby" -> "Sarah can be found at the lobby"
+
+DO NOT track current user position/state as knowledge - that changes constantly.
+DO track permanent facts learned from the user's actions.
+
+## PRESERVE SPECIFIC DETAILS
+Keep names, locations, numbers, and other specifics. Do NOT:
+- Abstract into general principles
+- Generate business insights
+- Make knowledge generic
+
+GOOD examples:
+- Fact: "John likes pizza" -> "John likes pizza"
+- Fact: "Alice works at Google" -> "Alice works at Google"
+
+BAD examples:
+- "John likes pizza" -> "Understanding dietary preferences helps..." (TOO ABSTRACT)
+- "User is at Room 203" -> "User is currently at Room 203" (EPHEMERAL STATE)
+
+## MERGE RULES (when comparing to existing observations):
+1. REDUNDANT: Same information worded differently → update existing
+2. CONTRADICTION: Opposite information about same topic → update with temporal markers showing change
+   Example: "Alex used to love pizza but now hates it" OR "Alex's pizza preference changed from love to hate"
+3. UPDATE: New state replacing old state → update showing the transition with "used to", "now", "changed from X to Y"
+
+## CRITICAL RULES:
+- NEVER merge facts about DIFFERENT people
+- NEVER merge unrelated topics (food preferences vs work vs hobbies)
+- When merging contradictions, the "text" field MUST capture BOTH states with temporal markers:
+  * Use "used to X, now Y" OR "changed from X to Y" OR "X but now Y"
+  * DO NOT just state the new fact - you MUST show the change
+- Keep observations focused on ONE specific topic per person
+- The "text" field MUST contain durable knowledge, not ephemeral state
+- Do NOT include "tags" in output - tags are handled automatically"""
+
+CONSOLIDATION_USER_PROMPT = """Analyze this new fact and consolidate into knowledge.
+{mission_section}
+NEW FACT: {fact_text}
+
+EXISTING OBSERVATIONS (JSON array with source memories and dates):
+{observations_text}
+
+Each observation includes:
+- id: unique identifier for updating
+- text: the observation content
+- proof_count: number of supporting memories
+- tags: visibility scope (handled automatically)
+- created_at/updated_at: when observation was created/modified
+- occurred_start/occurred_end: temporal range of source facts
+- source_memories: array of supporting facts with their text and dates
+
+Instructions:
+1. Extract DURABLE KNOWLEDGE from the new fact (not ephemeral state)
+2. Review source_memories in existing observations to understand evidence
+3. Check dates to detect contradictions or updates
+4. Compare with observations:
+   - Same topic → UPDATE with learning_id
+   - New topic → CREATE new observation
+   - Purely ephemeral → return []
+
+Output JSON array of actions (the "text" field should use markdown formatting for structure):
+[
+  {{"action": "update", "learning_id": "uuid-from-observations", "text": "## Updated Knowledge\n\n**Key point**: details here\n\n- Supporting detail 1\n- Supporting detail 2", "reason": "..."}},
+  {{"action": "create", "text": "## New Durable Knowledge\n\nDescription with **emphasis** and proper structure", "reason": "..."}}
+]
+
+Return [] if fact contains no durable knowledge.
+
+IMPORTANT: Format the "text" field with markdown for better readability:
+- Use headers, lists, bold/italic, tables where appropriate
+- CRITICAL: Add blank lines before and after block elements (tables, code blocks, lists)
+- Ensure proper spacing for markdown to render correctly"""

hindsight-api/hindsight_api/engine/cross_encoder.py

@@ -9,19 +9,35 @@ Configuration via environment variables - see hindsight_api.config for all env v
 import asyncio
 import logging
 import os
+import warnings
 from abc import ABC, abstractmethod
+from concurrent.futures import ThreadPoolExecutor
 
 import httpx
 
 from ..config import (
+    DEFAULT_LITELLM_API_BASE,
     DEFAULT_RERANKER_COHERE_MODEL,
+    DEFAULT_RERANKER_FLASHRANK_CACHE_DIR,
+    DEFAULT_RERANKER_FLASHRANK_MODEL,
+    DEFAULT_RERANKER_LITELLM_MODEL,
+    DEFAULT_RERANKER_LITELLM_SDK_MODEL,
+    DEFAULT_RERANKER_LOCAL_FORCE_CPU,
+    DEFAULT_RERANKER_LOCAL_MAX_CONCURRENT,
     DEFAULT_RERANKER_LOCAL_MODEL,
+    DEFAULT_RERANKER_LOCAL_TRUST_REMOTE_CODE,
     DEFAULT_RERANKER_PROVIDER,
     DEFAULT_RERANKER_TEI_BATCH_SIZE,
     DEFAULT_RERANKER_TEI_MAX_CONCURRENT,
-    ENV_COHERE_API_KEY,
+    ENV_RERANKER_COHERE_API_KEY,
     ENV_RERANKER_COHERE_MODEL,
+    ENV_RERANKER_FLASHRANK_CACHE_DIR,
+    ENV_RERANKER_FLASHRANK_MODEL,
+    ENV_RERANKER_LITELLM_SDK_API_KEY,
+    ENV_RERANKER_LOCAL_FORCE_CPU,
+    ENV_RERANKER_LOCAL_MAX_CONCURRENT,
     ENV_RERANKER_LOCAL_MODEL,
+    ENV_RERANKER_LOCAL_TRUST_REMOTE_CODE,
     ENV_RERANKER_PROVIDER,
     ENV_RERANKER_TEI_BATCH_SIZE,
     ENV_RERANKER_TEI_MAX_CONCURRENT,
@@ -78,25 +94,47 @@ class LocalSTCrossEncoder(CrossEncoderModel):
     - Fast inference (~80ms for 100 pairs on CPU)
     - Small model (80MB)
     - Trained for passage re-ranking
+
+    Uses a dedicated thread pool to limit concurrent CPU-bound work.
     """
 
-    def __init__(self, model_name: str | None = None):
+    # Shared executor across all instances (one model loaded anyway)
+    _executor: ThreadPoolExecutor | None = None
+    _max_concurrent: int = 4  # Limit concurrent CPU-bound reranking calls
+
+    def __init__(
+        self,
+        model_name: str | None = None,
+        max_concurrent: int = 4,
+        force_cpu: bool = False,
+        trust_remote_code: bool = False,
+    ):
         """
         Initialize local SentenceTransformers cross-encoder.
 
         Args:
             model_name: Name of the CrossEncoder model to use.
                        Default: cross-encoder/ms-marco-MiniLM-L-6-v2
+            max_concurrent: Maximum concurrent reranking calls (default: 2).
+                           Higher values may cause CPU thrashing under load.
+            force_cpu: Force CPU mode (avoids MPS/XPC issues on macOS in daemon mode).
+                      Default: False
+            trust_remote_code: Allow loading models with custom code (security risk).
+                              Required for some models like jina-reranker-v2-base-multilingual.
+                              Default: False (disabled for security)
         """
         self.model_name = model_name or DEFAULT_RERANKER_LOCAL_MODEL
+        self.force_cpu = force_cpu
+        self.trust_remote_code = trust_remote_code
         self._model = None
+        LocalSTCrossEncoder._max_concurrent = max_concurrent
 
     @property
     def provider_name(self) -> str:
         return "local"
 
     async def initialize(self) -> None:
-        """Load the cross-encoder model."""
+        """Load the cross-encoder model and initialize the executor."""
         if self._model is not None:
             return
 
@@ -109,13 +147,77 @@ class LocalSTCrossEncoder(CrossEncoderModel):
             )
 
         logger.info(f"Reranker: initializing local provider with model {self.model_name}")
-        self._model = CrossEncoder(self.model_name)
-        logger.info("Reranker: local provider initialized")
+
+        # Determine device based on hardware availability.
+        # We always set low_cpu_mem_usage=False to prevent lazy loading (meta tensors)
+        # which can cause issues when accelerate is installed but no GPU is available.
+        # Note: We do NOT use device_map because CrossEncoder internally calls .to(device)
+        # after loading, which conflicts with accelerate's device_map handling.
+        import torch
+
+        # Force CPU mode if configured (used in daemon mode to avoid MPS/XPC issues on macOS)
+        if self.force_cpu:
+            device = "cpu"
+            logger.info("Reranker: forcing CPU mode (HINDSIGHT_API_RERANKER_LOCAL_FORCE_CPU=1)")
+        else:
+            # Check for GPU (CUDA) or Apple Silicon (MPS)
+            # Wrap in try-except to gracefully handle any device detection issues
+            # (e.g., in CI environments or when PyTorch is built without GPU support)
+            device = "cpu"  # Default to CPU
+            try:
+                has_gpu = torch.cuda.is_available() or (
+                    hasattr(torch.backends, "mps") and torch.backends.mps.is_available()
+                )
+                if has_gpu:
+                    device = None  # Let sentence-transformers auto-detect GPU/MPS
+            except Exception as e:
+                logger.warning(f"Failed to detect GPU/MPS, falling back to CPU: {e}")
+
+        # Suppress verbose transformers warnings during model loading
+        # This suppresses the "UNEXPECTED" warnings from CrossEncoder which are harmless
+        # but look alarming to users (e.g., "embeddings.position_ids | UNEXPECTED")
+        with warnings.catch_warnings():
+            warnings.filterwarnings("ignore", category=UserWarning)
+            warnings.filterwarnings("ignore", message=".*was not found in model state dict.*")
+            warnings.filterwarnings("ignore", message=".*UNEXPECTED.*")
+
+            # Also suppress transformers library logging temporarily
+            transformers_logger = logging.getLogger("transformers")
+            original_level = transformers_logger.level
+            transformers_logger.setLevel(logging.ERROR)
+
+            try:
+                self._model = CrossEncoder(
+                    self.model_name,
+                    device=device,
+                    model_kwargs={"low_cpu_mem_usage": False},
+                    trust_remote_code=self.trust_remote_code,
+                )
+            finally:
+                # Restore original logging level
+                transformers_logger.setLevel(original_level)
+
+        # Initialize shared executor (limited workers naturally limits concurrency)
+        if LocalSTCrossEncoder._executor is None:
+            LocalSTCrossEncoder._executor = ThreadPoolExecutor(
+                max_workers=LocalSTCrossEncoder._max_concurrent,
+                thread_name_prefix="reranker",
+            )
+            logger.info(f"Reranker: local provider initialized (max_concurrent={LocalSTCrossEncoder._max_concurrent})")
+        else:
+            logger.info("Reranker: local provider initialized (using existing executor)")
+
+    def _predict_sync(self, pairs: list[tuple[str, str]]) -> list[float]:
+        """Synchronous prediction wrapper for thread pool execution."""
+        scores = self._model.predict(pairs, show_progress_bar=False)
+        return scores.tolist() if hasattr(scores, "tolist") else list(scores)
 
     async def predict(self, pairs: list[tuple[str, str]]) -> list[float]:
         """
         Score query-document pairs for relevance.
 
+        Uses a dedicated thread pool with limited workers to prevent CPU thrashing.
+
         Args:
             pairs: List of (query, document) tuples to score
 
@@ -125,10 +227,13 @@ class LocalSTCrossEncoder(CrossEncoderModel):
         if self._model is None:
             raise RuntimeError("Reranker not initialized. Call initialize() first.")
 
-        # Run CPU-bound inference in thread pool to avoid blocking event loop
+        # Use dedicated executor - limited workers naturally limits concurrency
         loop = asyncio.get_event_loop()
-        scores = await loop.run_in_executor(None, lambda: self._model.predict(pairs, show_progress_bar=False))
-        return scores.tolist() if hasattr(scores, "tolist") else list(scores)
+        return await loop.run_in_executor(
+            LocalSTCrossEncoder._executor,
+            self._predict_sync,
+            pairs,
+        )
 
 
 class RemoteTEICrossEncoder(CrossEncoderModel):
@@ -141,8 +246,13 @@ class RemoteTEICrossEncoder(CrossEncoderModel):
     Note: The TEI server must be running a cross-encoder/reranker model.
 
     Requests are made in parallel with configurable batch size and max concurrency (backpressure).
+    Uses a GLOBAL semaphore to limit concurrent requests across ALL recall operations.
     """
 
+    # Global semaphore shared across all instances and calls to prevent thundering herd
+    _global_semaphore: asyncio.Semaphore | None = None
+    _global_max_concurrent: int = DEFAULT_RERANKER_TEI_MAX_CONCURRENT
+
     def __init__(
         self,
         base_url: str,
@@ -159,7 +269,8 @@ class RemoteTEICrossEncoder(CrossEncoderModel):
             base_url: Base URL of the TEI server (e.g., "http://localhost:8080")
             timeout: Request timeout in seconds (default: 30.0)
             batch_size: Maximum batch size for rerank requests (default: 128)
-            max_concurrent: Maximum concurrent requests for backpressure (default: 8)
+            max_concurrent: Maximum concurrent requests for backpressure (default: 8).
+                           This is a GLOBAL limit across all parallel recall operations.
             max_retries: Maximum number of retries for failed requests (default: 3)
             retry_delay: Initial delay between retries in seconds, doubles each retry (default: 0.5)
         """
@@ -172,6 +283,14 @@ class RemoteTEICrossEncoder(CrossEncoderModel):
         self._async_client: httpx.AsyncClient | None = None
         self._model_id: str | None = None
 
+        # Update global semaphore if max_concurrent changed
+        if (
+            RemoteTEICrossEncoder._global_semaphore is None
+            or RemoteTEICrossEncoder._global_max_concurrent != max_concurrent
+        ):
+            RemoteTEICrossEncoder._global_max_concurrent = max_concurrent
+            RemoteTEICrossEncoder._global_semaphore = asyncio.Semaphore(max_concurrent)
+
     @property
     def provider_name(self) -> str:
         return "tei"
@@ -296,13 +415,13 @@ class RemoteTEICrossEncoder(CrossEncoderModel):
                 batch_texts = texts[i : i + self.batch_size]
                 tasks_info.append((query, batch_indices, batch_texts))
 
-        # Run all requests in parallel with semaphore for backpressure
+        # Run all requests in parallel with GLOBAL semaphore for backpressure
+        # This ensures max_concurrent is respected across ALL parallel recall operations
         all_scores = [0.0] * len(pairs)
-        semaphore = asyncio.Semaphore(self.max_concurrent)
+        semaphore = RemoteTEICrossEncoder._global_semaphore
 
         tasks = [
-            self._rerank_query_group(self._async_client, semaphore, query, texts)
-            for query, _, texts in tasks_info
+            self._rerank_query_group(self._async_client, semaphore, query, texts) for query, _, texts in tasks_info
         ]
         results = await asyncio.gather(*tasks)
 
@@ -343,6 +462,7 @@ class CohereCrossEncoder(CrossEncoderModel):
         self,
         api_key: str,
         model: str = DEFAULT_RERANKER_COHERE_MODEL,
+        base_url: str | None = None,
         timeout: float = 60.0,
     ):
         """
@@ -351,10 +471,12 @@ class CohereCrossEncoder(CrossEncoderModel):
         Args:
             api_key: Cohere API key
             model: Cohere rerank model name (default: rerank-english-v3.0)
+            base_url: Custom base URL for Cohere-compatible API (e.g., Azure-hosted endpoint)
             timeout: Request timeout in seconds (default: 60.0)
         """
         self.api_key = api_key
         self.model = model
+        self.base_url = base_url
         self.timeout = timeout
         self._client = None
 
@@ -372,8 +494,14 @@ class CohereCrossEncoder(CrossEncoderModel):
         except ImportError:
             raise ImportError("cohere is required for CohereCrossEncoder. Install it with: pip install cohere")
 
-        logger.info(f"Reranker: initializing Cohere provider with model {self.model}")
-        self._client = cohere.Client(api_key=self.api_key, timeout=self.timeout)
+        base_url_msg = f" at {self.base_url}" if self.base_url else ""
+        logger.info(f"Reranker: initializing Cohere provider with model {self.model}{base_url_msg}")
+
+        # Build client kwargs, only including base_url if set (for Azure or custom endpoints)
+        client_kwargs = {"api_key": self.api_key, "timeout": self.timeout}
+        if self.base_url:
+            client_kwargs["base_url"] = self.base_url
+        self._client = cohere.Client(**client_kwargs)
         logger.info("Reranker: Cohere provider initialized")
 
     async def predict(self, pairs: list[tuple[str, str]]) -> list[float]:
@@ -428,33 +556,463 @@ class CohereCrossEncoder(CrossEncoderModel):
         return all_scores
 
 
+class RRFPassthroughCrossEncoder(CrossEncoderModel):
+    """
+    Passthrough cross-encoder that preserves RRF scores without neural reranking.
+
+    This is useful for:
+    - Testing retrieval quality without reranking overhead
+    - Deployments where reranking latency is unacceptable
+    - Debugging to isolate retrieval vs reranking issues
+    """
+
+    def __init__(self):
+        """Initialize RRF passthrough cross-encoder."""
+        pass
+
+    @property
+    def provider_name(self) -> str:
+        return "rrf"
+
+    async def initialize(self) -> None:
+        """No initialization needed."""
+        logger.info("Reranker: RRF passthrough provider initialized (neural reranking disabled)")
+
+    async def predict(self, pairs: list[tuple[str, str]]) -> list[float]:
+        """
+        Return neutral scores - actual ranking uses RRF scores from retrieval.
+
+        Args:
+            pairs: List of (query, document) tuples (ignored)
+
+        Returns:
+            List of 0.5 scores (neutral, lets RRF scores dominate)
+        """
+        # Return neutral scores so RRF ranking is preserved
+        return [0.5] * len(pairs)
+
+
+class FlashRankCrossEncoder(CrossEncoderModel):
+    """
+    FlashRank cross-encoder implementation.
+
+    FlashRank is an ultra-lite reranking library that runs on CPU without
+    requiring PyTorch or Transformers. It's ideal for serverless deployments
+    with minimal cold-start overhead.
+
+    Available models:
+    - ms-marco-TinyBERT-L-2-v2: Fastest, ~4MB
+    - ms-marco-MiniLM-L-12-v2: Best quality, ~34MB (default)
+    - rank-T5-flan: Best zero-shot, ~110MB
+    - ms-marco-MultiBERT-L-12: Multi-lingual, ~150MB
+    """
+
+    # Shared executor for CPU-bound reranking
+    _executor: ThreadPoolExecutor | None = None
+    _max_concurrent: int = 4
+
+    def __init__(
+        self,
+        model_name: str | None = None,
+        cache_dir: str | None = None,
+        max_length: int = 512,
+        max_concurrent: int = 4,
+    ):
+        """
+        Initialize FlashRank cross-encoder.
+
+        Args:
+            model_name: FlashRank model name. Default: ms-marco-MiniLM-L-12-v2
+            cache_dir: Directory to cache downloaded models. Default: system cache
+            max_length: Maximum sequence length for reranking. Default: 512
+            max_concurrent: Maximum concurrent reranking calls. Default: 4
+        """
+        self.model_name = model_name or DEFAULT_RERANKER_FLASHRANK_MODEL
+        self.cache_dir = cache_dir or DEFAULT_RERANKER_FLASHRANK_CACHE_DIR
+        self.max_length = max_length
+        self._ranker = None
+        FlashRankCrossEncoder._max_concurrent = max_concurrent
+
+    @property
+    def provider_name(self) -> str:
+        return "flashrank"
+
+    async def initialize(self) -> None:
+        """Load the FlashRank model."""
+        if self._ranker is not None:
+            return
+
+        try:
+            from flashrank import Ranker
+        except ImportError:
+            raise ImportError("flashrank is required for FlashRankCrossEncoder. Install it with: pip install flashrank")
+
+        logger.info(f"Reranker: initializing FlashRank provider with model {self.model_name}")
+
+        # Initialize ranker with optional cache directory
+        ranker_kwargs = {"model_name": self.model_name, "max_length": self.max_length}
+        if self.cache_dir:
+            ranker_kwargs["cache_dir"] = self.cache_dir
+
+        self._ranker = Ranker(**ranker_kwargs)
+
+        # Initialize shared executor
+        if FlashRankCrossEncoder._executor is None:
+            FlashRankCrossEncoder._executor = ThreadPoolExecutor(
+                max_workers=FlashRankCrossEncoder._max_concurrent,
+                thread_name_prefix="flashrank",
+            )
+            logger.info(
+                f"Reranker: FlashRank provider initialized (max_concurrent={FlashRankCrossEncoder._max_concurrent})"
+            )
+        else:
+            logger.info("Reranker: FlashRank provider initialized (using existing executor)")
+
+    def _predict_sync(self, pairs: list[tuple[str, str]]) -> list[float]:
+        """Synchronous predict - processes each query group."""
+        from flashrank import RerankRequest
+
+        if not pairs:
+            return []
+
+        # Group pairs by query
+        query_groups: dict[str, list[tuple[int, str]]] = {}
+        for idx, (query, text) in enumerate(pairs):
+            if query not in query_groups:
+                query_groups[query] = []
+            query_groups[query].append((idx, text))
+
+        all_scores = [0.0] * len(pairs)
+
+        for query, indexed_texts in query_groups.items():
+            # Build passages list for FlashRank
+            passages = [{"id": i, "text": text} for i, (_, text) in enumerate(indexed_texts)]
+            global_indices = [idx for idx, _ in indexed_texts]
+
+            # Create rerank request
+            request = RerankRequest(query=query, passages=passages)
+            results = self._ranker.rerank(request)
+
+            # Map scores back to original positions
+            for result in results:
+                local_idx = result["id"]
+                score = result["score"]
+                global_idx = global_indices[local_idx]
+                all_scores[global_idx] = score
+
+        return all_scores
+
+    async def predict(self, pairs: list[tuple[str, str]]) -> list[float]:
+        """
+        Score query-document pairs using FlashRank.
+
+        Args:
+            pairs: List of (query, document) tuples to score
+
+        Returns:
+            List of relevance scores (higher = more relevant)
+        """
+        if self._ranker is None:
+            raise RuntimeError("Reranker not initialized. Call initialize() first.")
+
+        # Run in thread pool to avoid blocking event loop
+        loop = asyncio.get_event_loop()
+        return await loop.run_in_executor(FlashRankCrossEncoder._executor, self._predict_sync, pairs)
+
+
+class LiteLLMCrossEncoder(CrossEncoderModel):
+    """
+    LiteLLM cross-encoder implementation using LiteLLM proxy's /rerank endpoint.
+
+    LiteLLM provides a unified interface for multiple reranking providers via
+    the Cohere-compatible /rerank endpoint.
+    See: https://docs.litellm.ai/docs/rerank
+
+    Supported providers via LiteLLM:
+    - Cohere (rerank-english-v3.0, etc.) - prefix with cohere/
+    - Together AI - prefix with together_ai/
+    - Azure AI - prefix with azure_ai/
+    - Jina AI - prefix with jina_ai/
+    - AWS Bedrock - prefix with bedrock/
+    - Voyage AI - prefix with voyage/
+    """
+
+    def __init__(
+        self,
+        api_base: str = DEFAULT_LITELLM_API_BASE,
+        api_key: str | None = None,
+        model: str = DEFAULT_RERANKER_LITELLM_MODEL,
+        timeout: float = 60.0,
+    ):
+        """
+        Initialize LiteLLM cross-encoder client.
+
+        Args:
+            api_base: Base URL of the LiteLLM proxy (default: http://localhost:4000)
+            api_key: API key for the LiteLLM proxy (optional, depends on proxy config)
+            model: Reranking model name (default: cohere/rerank-english-v3.0)
+                   Use provider prefix (e.g., cohere/, together_ai/, voyage/)
+            timeout: Request timeout in seconds (default: 60.0)
+        """
+        self.api_base = api_base.rstrip("/")
+        self.api_key = api_key
+        self.model = model
+        self.timeout = timeout
+        self._async_client: httpx.AsyncClient | None = None
+
+    @property
+    def provider_name(self) -> str:
+        return "litellm"
+
+    async def initialize(self) -> None:
+        """Initialize the async HTTP client."""
+        if self._async_client is not None:
+            return
+
+        logger.info(f"Reranker: initializing LiteLLM provider at {self.api_base} with model {self.model}")
+
+        headers = {"Content-Type": "application/json"}
+        if self.api_key:
+            headers["Authorization"] = f"Bearer {self.api_key}"
+
+        self._async_client = httpx.AsyncClient(timeout=self.timeout, headers=headers)
+        logger.info("Reranker: LiteLLM provider initialized")
+
+    async def predict(self, pairs: list[tuple[str, str]]) -> list[float]:
+        """
+        Score query-document pairs using the LiteLLM proxy's /rerank endpoint.
+
+        Args:
+            pairs: List of (query, document) tuples to score
+
+        Returns:
+            List of relevance scores
+        """
+        if self._async_client is None:
+            raise RuntimeError("Reranker not initialized. Call initialize() first.")
+
+        if not pairs:
+            return []
+
+        # Group pairs by query (LiteLLM rerank expects one query with multiple documents)
+        query_groups: dict[str, list[tuple[int, str]]] = {}
+        for idx, (query, text) in enumerate(pairs):
+            if query not in query_groups:
+                query_groups[query] = []
+            query_groups[query].append((idx, text))
+
+        all_scores = [0.0] * len(pairs)
+
+        for query, indexed_texts in query_groups.items():
+            texts = [text for _, text in indexed_texts]
+            indices = [idx for idx, _ in indexed_texts]
+
+            # LiteLLM /rerank follows Cohere API format
+            response = await self._async_client.post(
+                f"{self.api_base}/rerank",
+                json={
+                    "model": self.model,
+                    "query": query,
+                    "documents": texts,
+                    "top_n": len(texts),  # Return all scores
+                },
+            )
+            response.raise_for_status()
+            result = response.json()
+
+            # Map scores back to original positions
+            # Response format: {"results": [{"index": 0, "relevance_score": 0.9}, ...]}
+            for item in result.get("results", []):
+                original_idx = item["index"]
+                score = item.get("relevance_score", item.get("score", 0.0))
+                all_scores[indices[original_idx]] = score
+
+        return all_scores
+
+
+class LiteLLMSDKCrossEncoder(CrossEncoderModel):
+    """
+    LiteLLM SDK cross-encoder for direct API integration.
+
+    Supports reranking via LiteLLM SDK without requiring a proxy server.
+    Supported providers: Cohere, DeepInfra, Together AI, HuggingFace, Jina AI, Voyage AI, AWS Bedrock.
+
+    Example model names:
+    - cohere/rerank-english-v3.0
+    - deepinfra/Qwen3-reranker-8B
+    - together_ai/Salesforce/Llama-Rank-V1
+    - huggingface/BAAI/bge-reranker-v2-m3
+    """
+
+    def __init__(
+        self,
+        api_key: str,
+        model: str = DEFAULT_RERANKER_LITELLM_SDK_MODEL,
+        api_base: str | None = None,
+        timeout: float = 60.0,
+    ):
+        """
+        Initialize LiteLLM SDK cross-encoder client.
+
+        Args:
+            api_key: API key for the reranking provider
+            model: Model name with provider prefix (e.g., "deepinfra/Qwen3-reranker-8B")
+            api_base: Custom base URL for API (optional)
+            timeout: Request timeout in seconds (default: 60.0)
+        """
+        self.api_key = api_key
+        self.model = model
+        self.api_base = api_base
+        self.timeout = timeout
+        self._initialized = False
+        self._litellm = None  # Will be set during initialization
+
+    @property
+    def provider_name(self) -> str:
+        return "litellm-sdk"
+
+    async def initialize(self) -> None:
+        """Initialize the LiteLLM SDK client."""
+        if self._initialized:
+            return
+
+        try:
+            import litellm
+
+            self._litellm = litellm  # Store reference
+        except ImportError:
+            raise ImportError("litellm is required for LiteLLMSDKCrossEncoder. Install it with: pip install litellm")
+
+        api_base_msg = f" at {self.api_base}" if self.api_base else ""
+        logger.info(f"Reranker: initializing LiteLLM SDK provider with model {self.model}{api_base_msg}")
+
+        self._initialized = True
+        logger.info("Reranker: LiteLLM SDK provider initialized")
+
+    async def predict(self, pairs: list[tuple[str, str]]) -> list[float]:
+        """
+        Score query-document pairs using the LiteLLM SDK.
+
+        Args:
+            pairs: List of (query, document) tuples to score
+
+        Returns:
+            List of relevance scores
+        """
+        if not self._initialized:
+            raise RuntimeError("Reranker not initialized. Call initialize() first.")
+
+        if not pairs:
+            return []
+
+        # Group pairs by query for efficient batching
+        # LiteLLM rerank expects one query with multiple documents
+        query_groups: dict[str, list[tuple[int, str]]] = {}
+        for idx, (query, text) in enumerate(pairs):
+            if query not in query_groups:
+                query_groups[query] = []
+            query_groups[query].append((idx, text))
+
+        all_scores = [0.0] * len(pairs)
+
+        for query, indexed_texts in query_groups.items():
+            texts = [text for _, text in indexed_texts]
+            indices = [idx for idx, _ in indexed_texts]
+
+            # Build kwargs for rerank call
+            rerank_kwargs = {
+                "model": self.model,
+                "query": query,
+                "documents": texts,
+                "api_key": self.api_key,
+            }
+            if self.api_base:
+                rerank_kwargs["api_base"] = self.api_base
+
+            response = await self._litellm.arerank(**rerank_kwargs)
+
+            # Map scores back to original positions
+            # Response format: RerankResponse with results list
+            # Each result is a TypedDict with "index" and "relevance_score"
+            if hasattr(response, "results") and response.results:
+                for result in response.results:
+                    # Results are TypedDicts, use dict-style access
+                    original_idx = result["index"]
+                    score = result.get("relevance_score", result.get("score", 0.0))
+                    all_scores[indices[original_idx]] = score
+            elif isinstance(response, list):
+                # Direct list of scores (unlikely but defensive)
+                for i, score in enumerate(response):
+                    all_scores[indices[i]] = score
+            else:
+                logger.warning(f"Unexpected response format from LiteLLM rerank: {type(response)}")
+
+        return all_scores
+
+
 def create_cross_encoder_from_env() -> CrossEncoderModel:
     """
-    Create a CrossEncoderModel instance based on environment variables.
+    Create a CrossEncoderModel instance based on configuration.
 
-    See hindsight_api.config for environment variable names and defaults.
+    Reads configuration via get_config() to ensure consistency across the codebase.
 
     Returns:
         Configured CrossEncoderModel instance
     """
-    provider = os.environ.get(ENV_RERANKER_PROVIDER, DEFAULT_RERANKER_PROVIDER).lower()
+    from ..config import get_config
+
+    config = get_config()
+    provider = config.reranker_provider.lower()
 
     if provider == "tei":
-        url = os.environ.get(ENV_RERANKER_TEI_URL)
+        url = config.reranker_tei_url
         if not url:
             raise ValueError(f"{ENV_RERANKER_TEI_URL} is required when {ENV_RERANKER_PROVIDER} is 'tei'")
-        batch_size = int(os.environ.get(ENV_RERANKER_TEI_BATCH_SIZE, str(DEFAULT_RERANKER_TEI_BATCH_SIZE)))
-        max_concurrent = int(os.environ.get(ENV_RERANKER_TEI_MAX_CONCURRENT, str(DEFAULT_RERANKER_TEI_MAX_CONCURRENT)))
-        return RemoteTEICrossEncoder(base_url=url, batch_size=batch_size, max_concurrent=max_concurrent)
+        return RemoteTEICrossEncoder(
+            base_url=url,
+            batch_size=config.reranker_tei_batch_size,
+            max_concurrent=config.reranker_tei_max_concurrent,
+        )
     elif provider == "local":
-        model = os.environ.get(ENV_RERANKER_LOCAL_MODEL)
-        model_name = model or DEFAULT_RERANKER_LOCAL_MODEL
-        return LocalSTCrossEncoder(model_name=model_name)
+        return LocalSTCrossEncoder(
+            model_name=config.reranker_local_model,
+            max_concurrent=config.reranker_local_max_concurrent,
+            force_cpu=config.reranker_local_force_cpu,
+            trust_remote_code=config.reranker_local_trust_remote_code,
+        )
     elif provider == "cohere":
-        api_key = os.environ.get(ENV_COHERE_API_KEY)
+        api_key = config.reranker_cohere_api_key
         if not api_key:
-            raise ValueError(f"{ENV_COHERE_API_KEY} is required when {ENV_RERANKER_PROVIDER} is 'cohere'")
-        model = os.environ.get(ENV_RERANKER_COHERE_MODEL, DEFAULT_RERANKER_COHERE_MODEL)
-        return CohereCrossEncoder(api_key=api_key, model=model)
+            raise ValueError(f"{ENV_RERANKER_COHERE_API_KEY} is required when {ENV_RERANKER_PROVIDER} is 'cohere'")
+        return CohereCrossEncoder(
+            api_key=api_key,
+            model=config.reranker_cohere_model,
+            base_url=config.reranker_cohere_base_url,
+        )
+    elif provider == "flashrank":
+        model = os.environ.get(ENV_RERANKER_FLASHRANK_MODEL, DEFAULT_RERANKER_FLASHRANK_MODEL)
+        cache_dir = os.environ.get(ENV_RERANKER_FLASHRANK_CACHE_DIR, DEFAULT_RERANKER_FLASHRANK_CACHE_DIR)
+        return FlashRankCrossEncoder(model_name=model, cache_dir=cache_dir)
+    elif provider == "litellm":
+        return LiteLLMCrossEncoder(
+            api_base=config.reranker_litellm_api_base,
+            api_key=config.reranker_litellm_api_key,
+            model=config.reranker_litellm_model,
+        )
+    elif provider == "litellm-sdk":
+        api_key = config.reranker_litellm_sdk_api_key
+        if not api_key:
+            raise ValueError(
+                f"{ENV_RERANKER_LITELLM_SDK_API_KEY} is required when {ENV_RERANKER_PROVIDER} is 'litellm-sdk'"
+            )
+        return LiteLLMSDKCrossEncoder(
+            api_key=api_key,
+            model=config.reranker_litellm_sdk_model,
+            api_base=config.reranker_litellm_sdk_api_base,
+        )
+    elif provider == "rrf":
+        return RRFPassthroughCrossEncoder()
     else:
-        raise ValueError(f"Unknown reranker provider: {provider}. Supported: 'local', 'tei', 'cohere'")
+        raise ValueError(
+            f"Unknown reranker provider: {provider}. Supported: 'local', 'tei', 'cohere', 'flashrank', 'litellm', 'litellm-sdk', 'rrf'"
+        )

hindsight-api/hindsight_api/engine/db_budget.py

@@ -0,0 +1,284 @@
+"""
+Database connection budget management.
+
+Limits concurrent database connections per operation to prevent
+a single operation (e.g., recall with parallel queries) from
+exhausting the connection pool.
+"""
+
+import asyncio
+import logging
+import uuid
+from contextlib import asynccontextmanager
+from dataclasses import dataclass, field
+from typing import TYPE_CHECKING, AsyncIterator
+
+if TYPE_CHECKING:
+    import asyncpg
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class OperationBudget:
+    """
+    Tracks connection budget for a single operation.
+
+    Each operation gets a semaphore limiting its concurrent connections.
+    """
+
+    operation_id: str
+    max_connections: int
+    semaphore: asyncio.Semaphore = field(init=False)
+    active_count: int = field(default=0, init=False)
+
+    def __post_init__(self):
+        self.semaphore = asyncio.Semaphore(self.max_connections)
+
+
+class ConnectionBudgetManager:
+    """
+    Manages per-operation connection budgets.
+
+    Usage:
+        manager = ConnectionBudgetManager(default_budget=4)
+
+        # Start an operation
+        async with manager.operation(max_connections=2) as op:
+            # Acquire connections within the budget
+            async with op.acquire(pool) as conn:
+                await conn.fetch(...)
+
+            # Multiple connections respect the budget
+            async with op.acquire(pool) as conn1, op.acquire(pool) as conn2:
+                # At most 2 concurrent connections for this operation
+                ...
+    """
+
+    def __init__(self, default_budget: int = 4):
+        """
+        Initialize the budget manager.
+
+        Args:
+            default_budget: Default max connections per operation
+        """
+        self.default_budget = default_budget
+        self._operations: dict[str, OperationBudget] = {}
+        self._lock = asyncio.Lock()
+
+    @asynccontextmanager
+    async def operation(
+        self,
+        max_connections: int | None = None,
+        operation_id: str | None = None,
+    ) -> AsyncIterator["BudgetedOperation"]:
+        """
+        Create a budgeted operation context.
+
+        Args:
+            max_connections: Max concurrent connections for this operation.
+                           Defaults to manager's default_budget.
+            operation_id: Optional custom operation ID. Auto-generated if not provided.
+
+        Yields:
+            BudgetedOperation context for acquiring connections
+        """
+        op_id = operation_id or f"op-{uuid.uuid4().hex[:12]}"
+        budget = max_connections or self.default_budget
+
+        async with self._lock:
+            if op_id in self._operations:
+                raise ValueError(f"Operation {op_id} already exists")
+            self._operations[op_id] = OperationBudget(op_id, budget)
+
+        try:
+            yield BudgetedOperation(self, op_id)
+        finally:
+            async with self._lock:
+                self._operations.pop(op_id, None)
+
+    def _get_budget(self, operation_id: str) -> OperationBudget:
+        """Get budget for an operation (internal use)."""
+        budget = self._operations.get(operation_id)
+        if not budget:
+            raise ValueError(f"Operation {operation_id} not found")
+        return budget
+
+
+class BudgetedOperation:
+    """
+    A single operation with connection budget.
+
+    Provides methods to acquire connections within the budget.
+    """
+
+    def __init__(self, manager: ConnectionBudgetManager, operation_id: str):
+        self._manager = manager
+        self.operation_id = operation_id
+
+    @property
+    def budget(self) -> OperationBudget:
+        """Get the budget for this operation."""
+        return self._manager._get_budget(self.operation_id)
+
+    @asynccontextmanager
+    async def acquire(self, pool: "asyncpg.Pool") -> AsyncIterator["asyncpg.Connection"]:
+        """
+        Acquire a connection within the operation's budget.
+
+        Blocks if the operation has reached its connection limit.
+
+        Args:
+            pool: asyncpg connection pool
+
+        Yields:
+            Database connection
+        """
+        budget = self.budget
+        async with budget.semaphore:
+            budget.active_count += 1
+            conn = await pool.acquire()
+            try:
+                yield conn
+            finally:
+                budget.active_count -= 1
+                await pool.release(conn)
+
+    def wrap_pool(self, pool: "asyncpg.Pool") -> "BudgetedPool":
+        """
+        Wrap a pool with this operation's budget.
+
+        The returned BudgetedPool can be passed to functions expecting a pool,
+        and all acquire() calls will be limited by this operation's budget.
+
+        Args:
+            pool: asyncpg connection pool to wrap
+
+        Returns:
+            BudgetedPool that limits connections to this operation's budget
+        """
+        return BudgetedPool(pool, self)
+
+    async def acquire_many(
+        self,
+        pool: "asyncpg.Pool",
+        count: int,
+    ) -> AsyncIterator[list["asyncpg.Connection"]]:
+        """
+        Acquire multiple connections within the budget.
+
+        Note: This acquires connections sequentially to respect the budget.
+        For parallel acquisition, use multiple acquire() calls with asyncio.gather().
+
+        Args:
+            pool: asyncpg connection pool
+            count: Number of connections to acquire
+
+        Yields:
+            List of database connections
+        """
+        connections = []
+        try:
+            for _ in range(count):
+                conn = await pool.acquire()
+                connections.append(conn)
+            yield connections
+        finally:
+            for conn in connections:
+                await pool.release(conn)
+
+
+# Global default manager instance
+_default_manager: ConnectionBudgetManager | None = None
+
+
+def get_budget_manager(default_budget: int = 4) -> ConnectionBudgetManager:
+    """
+    Get or create the global budget manager.
+
+    Args:
+        default_budget: Default max connections per operation
+
+    Returns:
+        Global ConnectionBudgetManager instance
+    """
+    global _default_manager
+    if _default_manager is None:
+        _default_manager = ConnectionBudgetManager(default_budget=default_budget)
+    return _default_manager
+
+
+@asynccontextmanager
+async def budgeted_operation(
+    max_connections: int | None = None,
+    operation_id: str | None = None,
+    default_budget: int = 4,
+) -> AsyncIterator[BudgetedOperation]:
+    """
+    Convenience function to create a budgeted operation.
+
+    Args:
+        max_connections: Max concurrent connections for this operation
+        operation_id: Optional custom operation ID
+        default_budget: Default budget if manager not yet created
+
+    Yields:
+        BudgetedOperation context
+
+    Example:
+        async with budgeted_operation(max_connections=2) as op:
+            async with op.acquire(pool) as conn:
+                await conn.fetch(...)
+    """
+    manager = get_budget_manager(default_budget)
+    async with manager.operation(max_connections, operation_id) as op:
+        yield op
+
+
+class BudgetedPool:
+    """
+    A pool wrapper that limits concurrent connection acquisitions.
+
+    This can be passed to functions expecting a pool, and acquire()
+    calls will be limited by the budget semaphore.
+
+    Usage:
+        async with budgeted_operation(max_connections=4) as op:
+            budgeted_pool = op.wrap_pool(pool)
+            # Pass budgeted_pool to functions that expect a pool
+            await some_function(budgeted_pool, ...)
+    """
+
+    def __init__(self, pool: "asyncpg.Pool", operation: BudgetedOperation):
+        self._pool = pool
+        self._operation = operation
+
+    async def acquire(self) -> "asyncpg.Connection":
+        """
+        Acquire a connection within the budget.
+
+        Note: Caller must release the connection when done.
+        Prefer using as context manager via acquire_with_retry or op.acquire().
+        """
+        budget = self._operation.budget
+        await budget.semaphore.acquire()
+        budget.active_count += 1
+        try:
+            return await self._pool.acquire()
+        except Exception:
+            budget.active_count -= 1
+            budget.semaphore.release()
+            raise
+
+    async def release(self, conn: "asyncpg.Connection") -> None:
+        """Release a connection back to the pool."""
+        budget = self._operation.budget
+        try:
+            await self._pool.release(conn)
+        finally:
+            budget.active_count -= 1
+            budget.semaphore.release()
+
+    def __getattr__(self, name):
+        """Proxy other attributes to the underlying pool."""
+        return getattr(self._pool, name)

hindsight-api/hindsight_api/engine/db_utils.py

@@ -83,11 +83,22 @@ async def acquire_with_retry(pool: asyncpg.Pool, max_retries: int = DEFAULT_MAX_
     Yields:
         An asyncpg connection
     """
+    import time
+
+    start = time.time()
 
     async def acquire():
         return await pool.acquire()
 
     conn = await retry_with_backoff(acquire, max_retries=max_retries)
+    acquire_time = time.time() - start
+
+    # Log slow connection acquisitions (indicates pool contention)
+    if acquire_time > 0.05:  # 50ms threshold
+        pool_size = pool.get_size()
+        pool_free = pool.get_idle_size()
+        logger.warning(f"[DB POOL] Slow acquire: {acquire_time:.3f}s | size={pool_size}, idle={pool_free}")
+
     try:
         yield conn
     finally:

hindsight-api/hindsight_api/engine/directives/__init__.py

@@ -0,0 +1,5 @@
+"""Directives module for hard rules injected into prompts."""
+
+from .models import Directive
+
+__all__ = ["Directive"]

hindsight-api/hindsight_api/engine/directives/models.py

@@ -0,0 +1,37 @@
+"""Pydantic models for directives."""
+
+from datetime import datetime, timezone
+from uuid import UUID
+
+from pydantic import BaseModel, Field
+
+
+class Directive(BaseModel):
+    """A directive is a hard rule injected into prompts.
+
+    Directives are user-defined rules that guide agent behavior. Unlike mental models
+    which are automatically consolidated from memories, directives are explicit
+    instructions that are always included in relevant prompts.
+
+    Examples:
+    - "Always respond in formal English"
+    - "Never share personal data with third parties"
+    - "Prefer conservative investment recommendations"
+    """
+
+    id: UUID = Field(description="Unique identifier")
+    bank_id: str = Field(description="Bank this directive belongs to")
+    name: str = Field(description="Human-readable name")
+    content: str = Field(description="The directive text to inject into prompts")
+    priority: int = Field(default=0, description="Higher priority directives are injected first")
+    is_active: bool = Field(default=True, description="Whether this directive is currently active")
+    tags: list[str] = Field(default_factory=list, description="Tags for filtering")
+    created_at: datetime = Field(
+        default_factory=lambda: datetime.now(timezone.utc), description="When this directive was created"
+    )
+    updated_at: datetime = Field(
+        default_factory=lambda: datetime.now(timezone.utc), description="When this directive was last updated"
+    )
+
+    class Config:
+        from_attributes = True

hindsight-api/hindsight_api/engine/embeddings.py

@@ -11,19 +11,28 @@ Configuration via environment variables - see hindsight_api.config for all env v
 
 import logging
 import os
+import warnings
 from abc import ABC, abstractmethod
 
 import httpx
 
 from ..config import (
     DEFAULT_EMBEDDINGS_COHERE_MODEL,
+    DEFAULT_EMBEDDINGS_LITELLM_MODEL,
+    DEFAULT_EMBEDDINGS_LITELLM_SDK_MODEL,
+    DEFAULT_EMBEDDINGS_LOCAL_FORCE_CPU,
     DEFAULT_EMBEDDINGS_LOCAL_MODEL,
+    DEFAULT_EMBEDDINGS_LOCAL_TRUST_REMOTE_CODE,
     DEFAULT_EMBEDDINGS_OPENAI_MODEL,
     DEFAULT_EMBEDDINGS_PROVIDER,
-    ENV_COHERE_API_KEY,
-    ENV_EMBEDDINGS_COHERE_MODEL,
+    DEFAULT_LITELLM_API_BASE,
+    ENV_EMBEDDINGS_COHERE_API_KEY,
+    ENV_EMBEDDINGS_LITELLM_SDK_API_KEY,
+    ENV_EMBEDDINGS_LOCAL_FORCE_CPU,
     ENV_EMBEDDINGS_LOCAL_MODEL,
+    ENV_EMBEDDINGS_LOCAL_TRUST_REMOTE_CODE,
     ENV_EMBEDDINGS_OPENAI_API_KEY,
+    ENV_EMBEDDINGS_OPENAI_BASE_URL,
     ENV_EMBEDDINGS_OPENAI_MODEL,
     ENV_EMBEDDINGS_PROVIDER,
     ENV_EMBEDDINGS_TEI_URL,
@@ -85,15 +94,22 @@ class LocalSTEmbeddings(Embeddings):
     The embedding dimension is auto-detected from the model.
     """
 
-    def __init__(self, model_name: str | None = None):
+    def __init__(self, model_name: str | None = None, force_cpu: bool = False, trust_remote_code: bool = False):
         """
         Initialize local SentenceTransformers embeddings.
 
         Args:
             model_name: Name of the SentenceTransformer model to use.
                        Default: BAAI/bge-small-en-v1.5
+            force_cpu: Force CPU mode (avoids MPS/XPC issues on macOS in daemon mode).
+                      Default: False
+            trust_remote_code: Allow loading models with custom code (security risk).
+                              Required for some models with custom architectures.
+                              Default: False (disabled for security)
         """
         self.model_name = model_name or DEFAULT_EMBEDDINGS_LOCAL_MODEL
+        self.force_cpu = force_cpu
+        self.trust_remote_code = trust_remote_code
         self._model = None
         self._dimension: int | None = None
 
@@ -121,12 +137,53 @@ class LocalSTEmbeddings(Embeddings):
             )
 
         logger.info(f"Embeddings: initializing local provider with model {self.model_name}")
-        # Disable lazy loading (meta tensors) which causes issues with newer transformers/accelerate
-        # Setting low_cpu_mem_usage=False and device_map=None ensures tensors are fully materialized
-        self._model = SentenceTransformer(
-            self.model_name,
-            model_kwargs={"low_cpu_mem_usage": False, "device_map": None},
-        )
+
+        # Determine device based on hardware availability.
+        # We always set low_cpu_mem_usage=False to prevent lazy loading (meta tensors)
+        # which can cause issues when accelerate is installed but no GPU is available.
+        import torch
+
+        # Force CPU mode if configured (used in daemon mode to avoid MPS/XPC issues on macOS)
+        if self.force_cpu:
+            device = "cpu"
+            logger.info("Embeddings: forcing CPU mode")
+        else:
+            # Check for GPU (CUDA) or Apple Silicon (MPS)
+            # Wrap in try-except to gracefully handle any device detection issues
+            # (e.g., in CI environments or when PyTorch is built without GPU support)
+            device = "cpu"  # Default to CPU
+            try:
+                has_gpu = torch.cuda.is_available() or (
+                    hasattr(torch.backends, "mps") and torch.backends.mps.is_available()
+                )
+                if has_gpu:
+                    device = None  # Let sentence-transformers auto-detect GPU/MPS
+            except Exception as e:
+                logger.warning(f"Failed to detect GPU/MPS, falling back to CPU: {e}")
+
+        # Suppress verbose transformers warnings during model loading
+        # This suppresses the "UNEXPECTED" warnings from BertModel which are harmless
+        # but look alarming to users (e.g., "embeddings.position_ids | UNEXPECTED")
+        with warnings.catch_warnings():
+            warnings.filterwarnings("ignore", category=UserWarning)
+            warnings.filterwarnings("ignore", message=".*was not found in model state dict.*")
+            warnings.filterwarnings("ignore", message=".*UNEXPECTED.*")
+
+            # Also suppress transformers library logging temporarily
+            transformers_logger = logging.getLogger("transformers")
+            original_level = transformers_logger.level
+            transformers_logger.setLevel(logging.ERROR)
+
+            try:
+                self._model = SentenceTransformer(
+                    self.model_name,
+                    device=device,
+                    model_kwargs={"low_cpu_mem_usage": False},
+                    trust_remote_code=self.trust_remote_code,
+                )
+            finally:
+                # Restore original logging level
+                transformers_logger.setLevel(original_level)
 
         self._dimension = self._model.get_sentence_embedding_dimension()
         logger.info(f"Embeddings: local provider initialized (dim: {self._dimension})")
@@ -143,6 +200,7 @@ class LocalSTEmbeddings(Embeddings):
         """
         if self._model is None:
             raise RuntimeError("Embeddings not initialized. Call initialize() first.")
+
         embeddings = self._model.encode(texts, convert_to_numpy=True, show_progress_bar=False)
         return [emb.tolist() for emb in embeddings]
 
@@ -322,6 +380,7 @@ class OpenAIEmbeddings(Embeddings):
         self,
         api_key: str,
         model: str = DEFAULT_EMBEDDINGS_OPENAI_MODEL,
+        base_url: str | None = None,
         batch_size: int = 100,
         max_retries: int = 3,
     ):
@@ -331,11 +390,13 @@ class OpenAIEmbeddings(Embeddings):
         Args:
             api_key: OpenAI API key
             model: OpenAI embedding model name (default: text-embedding-3-small)
+            base_url: Custom base URL for OpenAI-compatible API (e.g., Azure OpenAI endpoint)
             batch_size: Maximum batch size for embedding requests (default: 100)
             max_retries: Maximum number of retries for failed requests (default: 3)
         """
         self.api_key = api_key
         self.model = model
+        self.base_url = base_url
         self.batch_size = batch_size
         self.max_retries = max_retries
         self._client = None
@@ -361,8 +422,14 @@ class OpenAIEmbeddings(Embeddings):
         except ImportError:
             raise ImportError("openai is required for OpenAIEmbeddings. Install it with: pip install openai")
 
-        logger.info(f"Embeddings: initializing OpenAI provider with model {self.model}")
-        self._client = OpenAI(api_key=self.api_key, max_retries=self.max_retries)
+        base_url_msg = f" at {self.base_url}" if self.base_url else ""
+        logger.info(f"Embeddings: initializing OpenAI provider with model {self.model}{base_url_msg}")
+
+        # Build client kwargs, only including base_url if set (for Azure or custom endpoints)
+        client_kwargs = {"api_key": self.api_key, "max_retries": self.max_retries}
+        if self.base_url:
+            client_kwargs["base_url"] = self.base_url
+        self._client = OpenAI(**client_kwargs)
 
         # Try to get dimension from known models, otherwise do a test embedding
         if self.model in self.MODEL_DIMENSIONS:
@@ -435,6 +502,7 @@ class CohereEmbeddings(Embeddings):
         self,
         api_key: str,
         model: str = DEFAULT_EMBEDDINGS_COHERE_MODEL,
+        base_url: str | None = None,
         batch_size: int = 96,
         timeout: float = 60.0,
         input_type: str = "search_document",
@@ -445,6 +513,7 @@ class CohereEmbeddings(Embeddings):
         Args:
             api_key: Cohere API key
             model: Cohere embedding model name (default: embed-english-v3.0)
+            base_url: Custom base URL for Cohere-compatible API (e.g., Azure-hosted endpoint)
             batch_size: Maximum batch size for embedding requests (default: 96, Cohere's limit)
             timeout: Request timeout in seconds (default: 60.0)
             input_type: Input type for embeddings (default: search_document).
@@ -452,6 +521,7 @@ class CohereEmbeddings(Embeddings):
         """
         self.api_key = api_key
         self.model = model
+        self.base_url = base_url
         self.batch_size = batch_size
         self.timeout = timeout
         self.input_type = input_type
@@ -478,8 +548,14 @@ class CohereEmbeddings(Embeddings):
         except ImportError:
             raise ImportError("cohere is required for CohereEmbeddings. Install it with: pip install cohere")
 
-        logger.info(f"Embeddings: initializing Cohere provider with model {self.model}")
-        self._client = cohere.Client(api_key=self.api_key, timeout=self.timeout)
+        base_url_msg = f" at {self.base_url}" if self.base_url else ""
+        logger.info(f"Embeddings: initializing Cohere provider with model {self.model}{base_url_msg}")
+
+        # Build client kwargs, only including base_url if set (for Azure or custom endpoints)
+        client_kwargs = {"api_key": self.api_key, "timeout": self.timeout}
+        if self.base_url:
+            client_kwargs["base_url"] = self.base_url
+        self._client = cohere.Client(**client_kwargs)
 
         # Try to get dimension from known models, otherwise do a test embedding
         if self.model in self.MODEL_DIMENSIONS:
@@ -491,7 +567,7 @@ class CohereEmbeddings(Embeddings):
                 model=self.model,
                 input_type=self.input_type,
             )
-            if response.embeddings:
+            if response.embeddings and isinstance(response.embeddings, list):
                 self._dimension = len(response.embeddings[0])
 
         logger.info(f"Embeddings: Cohere provider initialized (model: {self.model}, dim: {self._dimension})")
@@ -529,26 +605,290 @@ class CohereEmbeddings(Embeddings):
         return all_embeddings
 
 
+class LiteLLMEmbeddings(Embeddings):
+    """
+    LiteLLM embeddings implementation using LiteLLM proxy's /embeddings endpoint.
+
+    LiteLLM provides a unified interface for multiple embedding providers.
+    The proxy exposes an OpenAI-compatible /embeddings endpoint.
+    See: https://docs.litellm.ai/docs/embedding/supported_embedding
+
+    Supported providers via LiteLLM:
+    - OpenAI (text-embedding-3-small, text-embedding-ada-002, etc.)
+    - Cohere (embed-english-v3.0, etc.) - prefix with cohere/
+    - Vertex AI (textembedding-gecko, etc.) - prefix with vertex_ai/
+    - HuggingFace, Mistral, Voyage AI, etc.
+
+    The embedding dimension is auto-detected from the model at initialization.
+    """
+
+    def __init__(
+        self,
+        api_base: str = DEFAULT_LITELLM_API_BASE,
+        api_key: str | None = None,
+        model: str = DEFAULT_EMBEDDINGS_LITELLM_MODEL,
+        batch_size: int = 100,
+        timeout: float = 60.0,
+    ):
+        """
+        Initialize LiteLLM embeddings client.
+
+        Args:
+            api_base: Base URL of the LiteLLM proxy (default: http://localhost:4000)
+            api_key: API key for the LiteLLM proxy (optional, depends on proxy config)
+            model: Embedding model name (default: text-embedding-3-small)
+                   Use provider prefix for non-OpenAI models (e.g., cohere/embed-english-v3.0)
+            batch_size: Maximum batch size for embedding requests (default: 100)
+            timeout: Request timeout in seconds (default: 60.0)
+        """
+        self.api_base = api_base.rstrip("/")
+        self.api_key = api_key
+        self.model = model
+        self.batch_size = batch_size
+        self.timeout = timeout
+        self._client: httpx.Client | None = None
+        self._dimension: int | None = None
+
+    @property
+    def provider_name(self) -> str:
+        return "litellm"
+
+    @property
+    def dimension(self) -> int:
+        if self._dimension is None:
+            raise RuntimeError("Embeddings not initialized. Call initialize() first.")
+        return self._dimension
+
+    async def initialize(self) -> None:
+        """Initialize the HTTP client and detect embedding dimension."""
+        if self._client is not None:
+            return
+
+        logger.info(f"Embeddings: initializing LiteLLM provider at {self.api_base} with model {self.model}")
+
+        headers = {"Content-Type": "application/json"}
+        if self.api_key:
+            headers["Authorization"] = f"Bearer {self.api_key}"
+
+        self._client = httpx.Client(timeout=self.timeout, headers=headers)
+
+        # Do a test embedding to detect dimension
+        try:
+            response = self._client.post(
+                f"{self.api_base}/embeddings",
+                json={"model": self.model, "input": ["test"]},
+            )
+            response.raise_for_status()
+            result = response.json()
+            if result.get("data") and len(result["data"]) > 0:
+                self._dimension = len(result["data"][0]["embedding"])
+            logger.info(f"Embeddings: LiteLLM provider initialized (model: {self.model}, dim: {self._dimension})")
+        except httpx.HTTPError as e:
+            raise RuntimeError(f"Failed to connect to LiteLLM proxy at {self.api_base}: {e}")
+
+    def encode(self, texts: list[str]) -> list[list[float]]:
+        """
+        Generate embeddings using the LiteLLM proxy.
+
+        Args:
+            texts: List of text strings to encode
+
+        Returns:
+            List of embedding vectors
+        """
+        if self._client is None:
+            raise RuntimeError("Embeddings not initialized. Call initialize() first.")
+
+        if not texts:
+            return []
+
+        all_embeddings = []
+
+        # Process in batches
+        for i in range(0, len(texts), self.batch_size):
+            batch = texts[i : i + self.batch_size]
+
+            response = self._client.post(
+                f"{self.api_base}/embeddings",
+                json={"model": self.model, "input": batch},
+            )
+            response.raise_for_status()
+            result = response.json()
+
+            # Sort by index to ensure correct order
+            batch_embeddings = sorted(result["data"], key=lambda x: x["index"])
+            all_embeddings.extend([e["embedding"] for e in batch_embeddings])
+
+        return all_embeddings
+
+
+class LiteLLMSDKEmbeddings(Embeddings):
+    """
+    LiteLLM SDK embeddings for direct API integration.
+
+    Supports embeddings via LiteLLM SDK without requiring a proxy server.
+    Supported providers: Cohere, OpenAI, Azure OpenAI, HuggingFace, Voyage AI, Together AI, etc.
+
+    Example model names:
+    - cohere/embed-english-v3.0
+    - openai/text-embedding-3-small
+    - together_ai/togethercomputer/m2-bert-80M-8k-retrieval
+    - voyage/voyage-2
+    """
+
+    def __init__(
+        self,
+        api_key: str,
+        model: str = DEFAULT_EMBEDDINGS_LITELLM_SDK_MODEL,
+        api_base: str | None = None,
+        batch_size: int = 100,
+        timeout: float = 60.0,
+    ):
+        """
+        Initialize LiteLLM SDK embeddings client.
+
+        Args:
+            api_key: API key for the embedding provider
+            model: Model name with provider prefix (e.g., "cohere/embed-english-v3.0")
+            api_base: Custom base URL for API (optional)
+            batch_size: Maximum batch size for embedding requests (default: 100)
+            timeout: Request timeout in seconds (default: 60.0)
+        """
+        self.api_key = api_key
+        self.model = model
+        self.api_base = api_base
+        self.batch_size = batch_size
+        self.timeout = timeout
+        self._litellm = None  # Will be set during initialization
+        self._dimension: int | None = None
+
+    @property
+    def provider_name(self) -> str:
+        return "litellm-sdk"
+
+    @property
+    def dimension(self) -> int:
+        if self._dimension is None:
+            raise RuntimeError("Embeddings not initialized. Call initialize() first.")
+        return self._dimension
+
+    async def initialize(self) -> None:
+        """Initialize the LiteLLM SDK client and detect dimension."""
+        if self._litellm is not None:
+            return
+
+        try:
+            import litellm
+
+            self._litellm = litellm  # Store reference
+        except ImportError:
+            raise ImportError("litellm is required for LiteLLMSDKEmbeddings. Install it with: pip install litellm")
+
+        api_base_msg = f" at {self.api_base}" if self.api_base else ""
+        logger.info(f"Embeddings: initializing LiteLLM SDK provider with model {self.model}{api_base_msg}")
+
+        # Do a test embedding to detect dimension
+        try:
+            # Build kwargs for embedding call
+            embed_kwargs = {
+                "model": self.model,
+                "input": ["test"],
+                "api_key": self.api_key,
+            }
+            if self.api_base:
+                embed_kwargs["api_base"] = self.api_base
+
+            # Use async embedding method (standard in litellm)
+            response = await self._litellm.aembedding(**embed_kwargs)
+
+            # Extract dimension from response
+            if response.data and len(response.data) > 0:
+                self._dimension = len(response.data[0]["embedding"])
+            else:
+                raise RuntimeError(f"Unable to detect embedding dimension for model {self.model}")
+
+        except Exception as e:
+            raise RuntimeError(f"Failed to initialize LiteLLM SDK embeddings: {e}")
+
+        logger.info(f"Embeddings: LiteLLM SDK provider initialized (model: {self.model}, dim: {self._dimension})")
+
+    def encode(self, texts: list[str]) -> list[list[float]]:
+        """
+        Generate embeddings using the LiteLLM SDK.
+
+        Args:
+            texts: List of text strings to encode
+
+        Returns:
+            List of embedding vectors (one per input text)
+        """
+        if self._litellm is None:
+            raise RuntimeError("Embeddings not initialized. Call initialize() first.")
+
+        if not texts:
+            return []
+
+        all_embeddings = []
+
+        # Process in batches
+        for i in range(0, len(texts), self.batch_size):
+            batch = texts[i : i + self.batch_size]
+
+            try:
+                # Build kwargs for embedding call
+                embed_kwargs = {
+                    "model": self.model,
+                    "input": batch,
+                    "api_key": self.api_key,
+                }
+                if self.api_base:
+                    embed_kwargs["api_base"] = self.api_base
+
+                # Use sync embedding (litellm doesn't have async in thread-safe way)
+                response = self._litellm.embedding(**embed_kwargs)
+
+                # Extract embeddings from response
+                # Sort by index to ensure correct order
+                batch_embeddings = sorted(response.data, key=lambda x: x.get("index", 0))
+                all_embeddings.extend([e["embedding"] for e in batch_embeddings])
+
+            except Exception as e:
+                import traceback
+
+                logger.error(
+                    f"Error in LiteLLM embedding for batch starting at index {i}: {e}\n"
+                    f"Traceback: {traceback.format_exc()}"
+                )
+                raise
+
+        return all_embeddings
+
+
 def create_embeddings_from_env() -> Embeddings:
     """
-    Create an Embeddings instance based on environment variables.
+    Create an Embeddings instance based on configuration.
 
-    See hindsight_api.config for environment variable names and defaults.
+    Reads configuration via get_config() to ensure consistency across the codebase.
 
     Returns:
         Configured Embeddings instance
     """
-    provider = os.environ.get(ENV_EMBEDDINGS_PROVIDER, DEFAULT_EMBEDDINGS_PROVIDER).lower()
+    from ..config import get_config
+
+    config = get_config()
+    provider = config.embeddings_provider.lower()
 
     if provider == "tei":
-        url = os.environ.get(ENV_EMBEDDINGS_TEI_URL)
+        url = config.embeddings_tei_url
         if not url:
             raise ValueError(f"{ENV_EMBEDDINGS_TEI_URL} is required when {ENV_EMBEDDINGS_PROVIDER} is 'tei'")
         return RemoteTEIEmbeddings(base_url=url)
     elif provider == "local":
-        model = os.environ.get(ENV_EMBEDDINGS_LOCAL_MODEL)
-        model_name = model or DEFAULT_EMBEDDINGS_LOCAL_MODEL
-        return LocalSTEmbeddings(model_name=model_name)
+        return LocalSTEmbeddings(
+            model_name=config.embeddings_local_model,
+            force_cpu=config.embeddings_local_force_cpu,
+            trust_remote_code=config.embeddings_local_trust_remote_code,
+        )
     elif provider == "openai":
         # Use dedicated embeddings API key, or fall back to LLM API key
         api_key = os.environ.get(ENV_EMBEDDINGS_OPENAI_API_KEY) or os.environ.get(ENV_LLM_API_KEY)
@@ -558,12 +898,36 @@ def create_embeddings_from_env() -> Embeddings:
                 f"when {ENV_EMBEDDINGS_PROVIDER} is 'openai'"
             )
         model = os.environ.get(ENV_EMBEDDINGS_OPENAI_MODEL, DEFAULT_EMBEDDINGS_OPENAI_MODEL)
-        return OpenAIEmbeddings(api_key=api_key, model=model)
+        base_url = os.environ.get(ENV_EMBEDDINGS_OPENAI_BASE_URL) or None
+        return OpenAIEmbeddings(api_key=api_key, model=model, base_url=base_url)
     elif provider == "cohere":
-        api_key = os.environ.get(ENV_COHERE_API_KEY)
+        api_key = config.embeddings_cohere_api_key
         if not api_key:
-            raise ValueError(f"{ENV_COHERE_API_KEY} is required when {ENV_EMBEDDINGS_PROVIDER} is 'cohere'")
-        model = os.environ.get(ENV_EMBEDDINGS_COHERE_MODEL, DEFAULT_EMBEDDINGS_COHERE_MODEL)
-        return CohereEmbeddings(api_key=api_key, model=model)
+            raise ValueError(f"{ENV_EMBEDDINGS_COHERE_API_KEY} is required when {ENV_EMBEDDINGS_PROVIDER} is 'cohere'")
+        return CohereEmbeddings(
+            api_key=api_key,
+            model=config.embeddings_cohere_model,
+            base_url=config.embeddings_cohere_base_url,
+        )
+    elif provider == "litellm":
+        return LiteLLMEmbeddings(
+            api_base=config.embeddings_litellm_api_base,
+            api_key=config.embeddings_litellm_api_key,
+            model=config.embeddings_litellm_model,
+        )
+    elif provider == "litellm-sdk":
+        api_key = config.embeddings_litellm_sdk_api_key
+        if not api_key:
+            raise ValueError(
+                f"{ENV_EMBEDDINGS_LITELLM_SDK_API_KEY} is required when {ENV_EMBEDDINGS_PROVIDER} is 'litellm-sdk'"
+            )
+        return LiteLLMSDKEmbeddings(
+            api_key=api_key,
+            model=config.embeddings_litellm_sdk_model,
+            api_base=config.embeddings_litellm_sdk_api_base,
+        )
     else:
-        raise ValueError(f"Unknown embeddings provider: {provider}. Supported: 'local', 'tei', 'openai', 'cohere'")
+        raise ValueError(
+            f"Unknown embeddings provider: {provider}. "
+            f"Supported: 'local', 'tei', 'openai', 'cohere', 'litellm', 'litellm-sdk'"
+        )

hindsight-api/hindsight_api/engine/interface.py

@@ -48,6 +48,7 @@ class MemoryEngineInterface(ABC):
         contents: list[dict[str, Any]],
         *,
         request_context: "RequestContext",
+        document_tags: list[str] | None = None,
     ) -> dict[str, Any]:
         """
         Retain a batch of memory items.
@@ -55,8 +56,9 @@ class MemoryEngineInterface(ABC):
         Args:
             bank_id: The memory bank ID.
             contents: List of content dicts with 'content', optional 'event_date',
-                     'context', 'metadata', 'document_id'.
+                     'context', 'metadata', 'document_id', and per-item 'tags'.
             request_context: Request context for authentication.
+            document_tags: Optional tags applied to all items in the batch.
 
         Returns:
             Dict with processing results.
@@ -160,14 +162,14 @@ class MemoryEngineInterface(ABC):
         request_context: "RequestContext",
     ) -> dict[str, Any]:
         """
-        Get bank profile including disposition and background.
+        Get bank profile including disposition and mission.
 
         Args:
             bank_id: The memory bank ID.
             request_context: Request context for authentication.
 
         Returns:
-            Bank profile dict.
+            Bank profile dict with bank_id, name, disposition, and mission.
         """
         ...
 
@@ -190,25 +192,44 @@ class MemoryEngineInterface(ABC):
         ...
 
     @abstractmethod
-    async def merge_bank_background(
+    async def merge_bank_mission(
         self,
         bank_id: str,
         new_info: str,
         *,
-        update_disposition: bool = True,
         request_context: "RequestContext",
     ) -> dict[str, Any]:
         """
-        Merge new background information into bank profile.
+        Merge new mission information into bank profile.
 
         Args:
             bank_id: The memory bank ID.
-            new_info: New background information to merge.
-            update_disposition: Whether to infer disposition from background.
+            new_info: New mission information to merge.
             request_context: Request context for authentication.
 
         Returns:
-            Updated background info.
+            Updated mission info.
+        """
+        ...
+
+    @abstractmethod
+    async def set_bank_mission(
+        self,
+        bank_id: str,
+        mission: str,
+        *,
+        request_context: "RequestContext",
+    ) -> dict[str, Any]:
+        """
+        Set the bank's mission (replaces existing).
+
+        Args:
+            bank_id: The memory bank ID.
+            mission: The mission text.
+            request_context: Request context for authentication.
+
+        Returns:
+            Dict with bank_id and mission.
         """
         ...
 
@@ -406,61 +427,20 @@ class MemoryEngineInterface(ABC):
         bank_id: str,
         *,
         limit: int = 100,
+        offset: int = 0,
         request_context: "RequestContext",
-    ) -> list[dict[str, Any]]:
+    ) -> dict[str, Any]:
         """
-        List entities for a bank.
+        List entities for a bank with pagination.
 
         Args:
             bank_id: The memory bank ID.
             limit: Maximum results.
+            offset: Offset for pagination.
             request_context: Request context for authentication.
 
         Returns:
-            List of entity dicts.
-        """
-        ...
-
-    @abstractmethod
-    async def get_entity_observations(
-        self,
-        bank_id: str,
-        entity_id: str,
-        *,
-        limit: int = 10,
-        request_context: "RequestContext",
-    ) -> list[Any]:
-        """
-        Get observations for an entity.
-
-        Args:
-            bank_id: The memory bank ID.
-            entity_id: The entity ID.
-            limit: Maximum observations.
-            request_context: Request context for authentication.
-
-        Returns:
-            List of EntityObservation objects.
-        """
-        ...
-
-    @abstractmethod
-    async def regenerate_entity_observations(
-        self,
-        bank_id: str,
-        entity_id: str,
-        entity_name: str,
-        *,
-        request_context: "RequestContext",
-    ) -> None:
-        """
-        Regenerate observations for an entity.
-
-        Args:
-            bank_id: The memory bank ID.
-            entity_id: The entity ID.
-            entity_name: The entity's canonical name.
-            request_context: Request context for authentication.
+            Dict with items, total, limit, offset.
         """
         ...
 
@@ -516,7 +496,7 @@ class MemoryEngineInterface(ABC):
         bank_id: str,
         *,
         request_context: "RequestContext",
-    ) -> list[dict[str, Any]]:
+    ) -> dict[str, Any]:
         """
         List async operations for a bank.
 
@@ -525,7 +505,7 @@ class MemoryEngineInterface(ABC):
             request_context: Request context for authentication.
 
         Returns:
-            List of operation dicts with id, task_type, status, etc.
+            Dict with 'total' (int) and 'operations' (list of operation dicts).
         """
         ...
 
@@ -559,16 +539,16 @@ class MemoryEngineInterface(ABC):
         bank_id: str,
         *,
         name: str | None = None,
-        background: str | None = None,
+        mission: str | None = None,
         request_context: "RequestContext",
     ) -> dict[str, Any]:
         """
-        Update bank name and/or background.
+        Update bank name and/or mission.
 
         Args:
             bank_id: The memory bank ID.
             name: New bank name (optional).
-            background: New background text (optional, replaces existing).
+            mission: New mission text (optional, replaces existing).
             request_context: Request context for authentication.
 
         Returns:
@@ -583,6 +563,7 @@ class MemoryEngineInterface(ABC):
         contents: list[dict[str, Any]],
         *,
         request_context: "RequestContext",
+        document_tags: list[str] | None = None,
     ) -> dict[str, Any]:
         """
         Submit a batch retain operation to run asynchronously.
@@ -591,6 +572,7 @@ class MemoryEngineInterface(ABC):
             bank_id: The memory bank ID.
             contents: List of content dicts to retain.
             request_context: Request context for authentication.
+            document_tags: Optional tags applied to all items in the async batch.
 
         Returns:
             Dict with operation_id and items_count.

hindsight-api/hindsight_api/engine/llm_interface.py

@@ -0,0 +1,207 @@
+"""
+Abstract interface for LLM providers.
+
+This module defines the interface that all LLM providers must implement,
+enabling support for multiple LLM backends (OpenAI, Anthropic, Gemini, Codex, etc.)
+"""
+
+from abc import ABC, abstractmethod
+from typing import Any
+
+from .response_models import LLMToolCallResult, TokenUsage
+
+
+class LLMInterface(ABC):
+    """
+    Abstract interface for LLM providers.
+
+    All LLM provider implementations must inherit from this class and implement
+    the required methods.
+    """
+
+    def __init__(
+        self,
+        provider: str,
+        api_key: str,
+        base_url: str,
+        model: str,
+        reasoning_effort: str = "low",
+        **kwargs: Any,
+    ):
+        """
+        Initialize LLM provider.
+
+        Args:
+            provider: Provider name (e.g., "openai", "codex", "anthropic", "gemini").
+            api_key: API key or authentication token.
+            base_url: Base URL for the API.
+            model: Model name.
+            reasoning_effort: Reasoning effort level for supported providers.
+            **kwargs: Additional provider-specific parameters.
+        """
+        self.provider = provider.lower()
+        self.api_key = api_key
+        self.base_url = base_url
+        self.model = model
+        self.reasoning_effort = reasoning_effort
+
+    @abstractmethod
+    async def verify_connection(self) -> None:
+        """
+        Verify that the LLM provider is configured correctly by making a simple test call.
+
+        Raises:
+            RuntimeError: If the connection test fails.
+        """
+        pass
+
+    @abstractmethod
+    async def call(
+        self,
+        messages: list[dict[str, str]],
+        response_format: Any | None = None,
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "memory",
+        max_retries: int = 10,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 60.0,
+        skip_validation: bool = False,
+        strict_schema: bool = False,
+        return_usage: bool = False,
+    ) -> Any:
+        """
+        Make an LLM API call with retry logic.
+
+        Args:
+            messages: List of message dicts with 'role' and 'content'.
+            response_format: Optional Pydantic model for structured output.
+            max_completion_tokens: Maximum tokens in response.
+            temperature: Sampling temperature (0.0-2.0).
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            skip_validation: Return raw JSON without Pydantic validation.
+            strict_schema: Use strict JSON schema enforcement (OpenAI only).
+            return_usage: If True, return tuple (result, TokenUsage) instead of just result.
+
+        Returns:
+            If return_usage=False: Parsed response if response_format is provided, otherwise text content.
+            If return_usage=True: Tuple of (result, TokenUsage) with token counts.
+
+        Raises:
+            OutputTooLongError: If output exceeds token limits.
+            Exception: Re-raises API errors after retries exhausted.
+        """
+        pass
+
+    @abstractmethod
+    async def call_with_tools(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "tools",
+        max_retries: int = 5,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 30.0,
+        tool_choice: str | dict[str, Any] = "auto",
+    ) -> LLMToolCallResult:
+        """
+        Make an LLM API call with tool/function calling support.
+
+        Args:
+            messages: List of message dicts. Can include tool results with role='tool'.
+            tools: List of tool definitions in OpenAI format.
+            max_completion_tokens: Maximum tokens in response.
+            temperature: Sampling temperature (0.0-2.0).
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            tool_choice: How to choose tools - "auto", "none", "required", or specific function.
+
+        Returns:
+            LLMToolCallResult with content and/or tool_calls.
+        """
+        pass
+
+    async def supports_batch_api(self) -> bool:
+        """
+        Check if this provider supports batch API operations.
+
+        Returns:
+            True if provider supports submit_batch/get_batch_status/retrieve_batch_results
+        """
+        return False
+
+    async def submit_batch(
+        self,
+        requests: list[dict[str, Any]],
+        endpoint: str = "/v1/chat/completions",
+        completion_window: str = "24h",
+    ) -> dict[str, Any]:
+        """
+        Submit a batch of requests to the provider's batch API.
+
+        Args:
+            requests: List of request dicts in JSONL format (custom_id, method, url, body)
+            endpoint: API endpoint for the batch (e.g., "/v1/chat/completions")
+            completion_window: Completion window (e.g., "24h")
+
+        Returns:
+            Dict with batch metadata: {"batch_id": str, "status": str, ...}
+
+        Raises:
+            NotImplementedError: If provider doesn't support batch API
+        """
+        raise NotImplementedError(f"Batch API not supported for provider: {self.provider}")
+
+    async def get_batch_status(self, batch_id: str) -> dict[str, Any]:
+        """
+        Get the status of a batch job.
+
+        Args:
+            batch_id: Batch identifier returned from submit_batch
+
+        Returns:
+            Dict with status info: {"batch_id": str, "status": str, "completed_at": str, ...}
+
+        Raises:
+            NotImplementedError: If provider doesn't support batch API
+        """
+        raise NotImplementedError(f"Batch API not supported for provider: {self.provider}")
+
+    async def retrieve_batch_results(self, batch_id: str) -> list[dict[str, Any]]:
+        """
+        Retrieve completed batch results.
+
+        Args:
+            batch_id: Batch identifier returned from submit_batch
+
+        Returns:
+            List of result dicts (one per request, matched by custom_id)
+
+        Raises:
+            NotImplementedError: If provider doesn't support batch API
+        """
+        raise NotImplementedError(f"Batch API not supported for provider: {self.provider}")
+
+    @abstractmethod
+    async def cleanup(self) -> None:
+        """Clean up resources (close connections, etc.)."""
+        pass
+
+
+class OutputTooLongError(Exception):
+    """
+    Bridge exception raised when LLM output exceeds token limits.
+
+    This wraps provider-specific errors (e.g., OpenAI's LengthFinishReasonError)
+    to allow callers to handle output length issues without depending on
+    provider-specific implementations.
+    """
+
+    pass

hindsight-api/hindsight_api/engine/mental_models/__init__.py

@@ -0,0 +1,14 @@
+"""
+Mental models module for Hindsight.
+
+Mental models contain directives - hard rules that are injected into reflect prompts.
+Directives are user-defined and their observations are user-provided (not LLM-generated).
+
+Other types of consolidated knowledge are handled by:
+- Learnings: Automatic bottom-up consolidation from facts
+- Pinned Reflections: User-curated living documents
+"""
+
+from .models import MentalModel, MentalModelSubtype
+
+__all__ = ["MentalModel", "MentalModelSubtype"]

hindsight-api/hindsight_api/engine/mental_models/models.py

@@ -0,0 +1,53 @@
+"""
+Pydantic models for mental models.
+"""
+
+from datetime import datetime, timezone
+from enum import Enum
+
+from pydantic import BaseModel, Field
+
+
+class MentalModelSubtype(str, Enum):
+    """Subtype of mental model.
+
+    Currently only DIRECTIVE is supported. Other types of consolidated knowledge
+    are handled by:
+    - Learnings: Automatic bottom-up consolidation from facts
+    - Pinned Reflections: User-curated living documents
+    """
+
+    DIRECTIVE = "directive"  # User-defined hard rules, observations user-provided
+
+
+class MentalModel(BaseModel):
+    """
+    A mental model representing synthesized understanding.
+
+    Mental models are the agent's consolidated knowledge. Unlike raw facts,
+    mental models provide:
+    - A one-liner description for quick scanning/retrieval
+    - A full summary for deep understanding
+    - Links to related mental models
+    """
+
+    id: str = Field(description="Unique identifier within the bank")
+    bank_id: str = Field(description="Bank this mental model belongs to")
+    subtype: MentalModelSubtype = Field(description="How this model was created")
+    name: str = Field(description="Human-readable name")
+    description: str = Field(description="One-liner for quick scanning and retrieval matching")
+    summary: str | None = Field(default=None, description="Full synthesized understanding")
+
+    # References
+    entity_id: str | None = Field(default=None, description="Reference to entities table when type=entity")
+    source_facts: list[str] = Field(default_factory=list, description="Fact IDs used to generate summary")
+    links: list[str] = Field(default_factory=list, description="Related mental model IDs")
+
+    # Tags for scoped visibility (similar to document tags)
+    tags: list[str] = Field(default_factory=list, description="Tags for scoped visibility filtering")
+
+    # Timestamps
+    last_updated: datetime | None = Field(default=None, description="When summary was last regenerated")
+    created_at: datetime = Field(
+        default_factory=lambda: datetime.now(timezone.utc), description="When this model was created"
+    )

hindsight-api/hindsight_api/engine/operation_metadata.py

@@ -0,0 +1,69 @@
+"""
+Typed metadata models for async operations.
+
+These dataclasses define the structure of result_metadata for different operation types.
+The metadata is exposed in the API for debugging purposes and may change without notice.
+"""
+
+from dataclasses import asdict, dataclass
+from typing import Any
+
+
+@dataclass
+class BatchRetainParentMetadata:
+    """Metadata for parent batch_retain operations (when split into sub-batches)."""
+
+    items_count: int
+    total_tokens: int
+    num_sub_batches: int
+    is_parent: bool = True
+
+    def to_dict(self) -> dict[str, Any]:
+        """Convert to dict for JSON serialization."""
+        return asdict(self)
+
+
+@dataclass
+class BatchRetainChildMetadata:
+    """Metadata for child batch_retain operations (individual sub-batches)."""
+
+    items_count: int
+    parent_operation_id: str
+    sub_batch_index: int
+    total_sub_batches: int
+
+    def to_dict(self) -> dict[str, Any]:
+        """Convert to dict for JSON serialization."""
+        return asdict(self)
+
+
+@dataclass
+class RetainMetadata:
+    """Metadata for regular retain operations (non-batched, deprecated async path)."""
+
+    items_count: int
+
+    def to_dict(self) -> dict[str, Any]:
+        """Convert to dict for JSON serialization."""
+        return asdict(self)
+
+
+@dataclass
+class ConsolidationMetadata:
+    """Metadata for consolidation operations."""
+
+    # Currently empty, but structure for future fields
+    def to_dict(self) -> dict[str, Any]:
+        """Convert to dict for JSON serialization."""
+        return asdict(self)
+
+
+@dataclass
+class RefreshMentalModelMetadata:
+    """Metadata for mental model refresh operations."""
+
+    mental_model_id: str
+
+    def to_dict(self) -> dict[str, Any]:
+        """Convert to dict for JSON serialization."""
+        return asdict(self)

hindsight-api/hindsight_api/engine/parsers/__init__.py

@@ -0,0 +1,62 @@
+"""File parser implementations."""
+
+from .base import FileParser, UnsupportedFileTypeError
+from .iris import IrisParser
+from .markitdown import MarkitdownParser
+
+__all__ = ["FileParser", "UnsupportedFileTypeError", "IrisParser", "MarkitdownParser", "FileParserRegistry"]
+
+
+class FileParserRegistry:
+    """Registry for file parsers with auto-detection."""
+
+    def __init__(self):
+        """Initialize empty parser registry."""
+        self._parsers: dict[str, FileParser] = {}
+
+    def register(self, parser: FileParser):
+        """
+        Register a parser.
+
+        Args:
+            parser: FileParser instance
+        """
+        self._parsers[parser.name()] = parser
+
+    def get_parser(
+        self,
+        name: str | None,
+        filename: str,
+        content_type: str | None = None,
+    ) -> FileParser:
+        """
+        Get parser by name or auto-detect.
+
+        Args:
+            name: Parser name (e.g., "markitdown") or None for auto-detect
+            filename: File name for auto-detection
+            content_type: MIME type (optional)
+
+        Returns:
+            FileParser instance
+
+        Raises:
+            ValueError: If no suitable parser found
+        """
+        if name:
+            # Explicit parser requested — return it directly, let the parser
+            # raise UnsupportedFileTypeError from convert() if needed
+            if name not in self._parsers:
+                raise ValueError(f"Parser '{name}' not found. Available: {list(self._parsers.keys())}")
+            return self._parsers[name]
+
+        # Auto-detect parser
+        for parser in self._parsers.values():
+            if parser.supports(filename, content_type):
+                return parser
+
+        raise ValueError(f"No parser found for {filename}. Available parsers: {list(self._parsers.keys())}")
+
+    def list_parsers(self) -> list[str]:
+        """Get list of registered parser names."""
+        return list(self._parsers.keys())

hindsight-api/hindsight_api/engine/parsers/base.py

@@ -0,0 +1,58 @@
+"""Abstract base class for file parsers."""
+
+from abc import ABC, abstractmethod
+
+
+class UnsupportedFileTypeError(Exception):
+    """Raised by a parser when it does not support the given file type."""
+
+    pass
+
+
+class FileParser(ABC):
+    """Abstract base for file to markdown parsers."""
+
+    @abstractmethod
+    async def convert(self, file_data: bytes, filename: str) -> str:
+        """
+        Parse file to markdown.
+
+        Args:
+            file_data: Raw file bytes
+            filename: Original filename (used for format detection)
+
+        Returns:
+            Markdown content as string
+
+        Raises:
+            UnsupportedFileTypeError: If the file type is not supported by this parser
+            RuntimeError: If parsing fails for another reason
+        """
+        pass
+
+    def supports(self, filename: str, content_type: str | None = None) -> bool:
+        """
+        Check if parser supports this file type.
+
+        Override this for local/static extension-based filtering.
+        Parsers that delegate to a remote service should leave this as True
+        and raise UnsupportedFileTypeError from convert() instead.
+
+        Args:
+            filename: File name (used for extension check)
+            content_type: MIME type (optional)
+
+        Returns:
+            True if this parser can handle the file (default: True)
+        """
+        return True
+
+    @abstractmethod
+    def name(self) -> str:
+        """
+        Get parser name.
+
+        Returns:
+            Parser name (e.g., "markitdown")
+        """
+        pass

hindsight-api/hindsight_api/engine/parsers/iris.py

@@ -0,0 +1,137 @@
+"""Iris parser implementation using the Vectorize Iris HTTP API."""
+
+import asyncio
+import logging
+import mimetypes
+import time
+
+import httpx
+
+from .base import FileParser, UnsupportedFileTypeError
+
+logger = logging.getLogger(__name__)
+
+_IRIS_BASE_URL = "https://api.vectorize.io/v1"
+_DEFAULT_POLL_INTERVAL = 2.0  # seconds
+_DEFAULT_TIMEOUT = 300.0  # seconds
+
+
+class IrisParser(FileParser):
+    """
+    Iris file parser using the Vectorize Iris cloud extraction service.
+
+    Uploads files to the Vectorize Iris API, starts an extraction job,
+    and polls until the text is ready. The API determines which file types
+    are supported — UnsupportedFileTypeError is raised if the file is rejected.
+
+    Authentication:
+        Requires HINDSIGHT_API_FILE_PARSER_IRIS_TOKEN and
+        HINDSIGHT_API_FILE_PARSER_IRIS_ORG_ID environment variables,
+        or pass them explicitly via the constructor.
+    """
+
+    def __init__(
+        self,
+        token: str,
+        org_id: str,
+        poll_interval: float = _DEFAULT_POLL_INTERVAL,
+        timeout: float = _DEFAULT_TIMEOUT,
+    ):
+        """
+        Initialize iris parser.
+
+        Args:
+            token: Vectorize API token
+            org_id: Vectorize organization ID
+            poll_interval: Seconds between status poll requests (default: 2)
+            timeout: Maximum seconds to wait for extraction (default: 300)
+        """
+        self._token = token
+        self._org_id = org_id
+        self._poll_interval = poll_interval
+        self._timeout = timeout
+        self._auth_headers = {"Authorization": f"Bearer {token}"}
+
+    async def convert(self, file_data: bytes, filename: str) -> str:
+        """
+        Parse file to text using the Vectorize Iris API.
+
+        Raises:
+            UnsupportedFileTypeError: If the Iris API rejects the file type (4xx)
+            RuntimeError: If extraction fails for another reason
+        """
+        content_type = mimetypes.guess_type(filename)[0] or "application/octet-stream"
+
+        async with httpx.AsyncClient() as client:
+            # Step 1: Request a presigned upload URL
+            init_resp = await client.post(
+                f"{_IRIS_BASE_URL}/org/{self._org_id}/files",
+                headers=self._auth_headers,
+                json={"name": filename, "contentType": content_type},
+            )
+            _raise_for_status(init_resp, filename, "file upload init")
+            init_data = init_resp.json()
+            file_id: str = init_data["fileId"]
+            upload_url: str = init_data["uploadUrl"]
+
+            # Step 2: Upload the file bytes to the presigned URL (no auth header)
+            upload_resp = await client.put(
+                upload_url,
+                content=file_data,
+                headers={"Content-Type": content_type},
+            )
+            _raise_for_status(upload_resp, filename, "file upload")
+
+            # Step 3: Start extraction
+            extract_resp = await client.post(
+                f"{_IRIS_BASE_URL}/org/{self._org_id}/extraction",
+                headers=self._auth_headers,
+                json={"fileId": file_id},
+            )
+            _raise_for_status(extract_resp, filename, "start extraction")
+            extraction_id: str = extract_resp.json()["extractionId"]
+
+            # Step 4: Poll until ready or timeout
+            deadline = time.monotonic() + self._timeout
+            while True:
+                status_resp = await client.get(
+                    f"{_IRIS_BASE_URL}/org/{self._org_id}/extraction/{extraction_id}",
+                    headers=self._auth_headers,
+                )
+                _raise_for_status(status_resp, filename, "poll extraction status")
+                status_data = status_resp.json()
+
+                if status_data.get("ready"):
+                    data = status_data.get("data", {})
+                    if not data.get("success"):
+                        error = data.get("error", "unknown error")
+                        raise RuntimeError(f"Iris extraction failed for '{filename}': {error}")
+                    text = data.get("text")
+                    if not text:
+                        raise RuntimeError(f"No content extracted from '{filename}'")
+                    return text
+
+                if time.monotonic() >= deadline:
+                    raise RuntimeError(f"Iris extraction timed out after {self._timeout}s for '{filename}'")
+
+                await asyncio.sleep(self._poll_interval)
+
+    def name(self) -> str:
+        """Get parser name."""
+        return "iris"
+
+
+def _raise_for_status(response: httpx.Response, filename: str, step: str) -> None:
+    """
+    Raise an appropriate error including the response body on HTTP errors.
+
+    Raises UnsupportedFileTypeError for 4xx responses (file rejected by the API),
+    RuntimeError for other HTTP errors.
+    """
+    if not response.is_error:
+        return
+    body = response.text or "<empty>"
+    msg = f"Iris API error during {step} for '{filename}': {response.status_code} {response.reason_phrase} — {body}"
+    if response.is_client_error:
+        raise UnsupportedFileTypeError(msg)
+    raise RuntimeError(msg)

hindsight-api/hindsight_api/engine/parsers/markitdown.py

@@ -0,0 +1,109 @@
+"""Markitdown parser implementation."""
+
+import asyncio
+import logging
+import tempfile
+from pathlib import Path
+
+from .base import FileParser
+
+logger = logging.getLogger(__name__)
+
+
+class MarkitdownParser(FileParser):
+    """
+    Markitdown file parser.
+
+    Uses Microsoft's markitdown library to convert various file formats
+    to markdown including PDF, Office docs, images (via OCR), audio, HTML.
+
+    Supported formats:
+    - PDF (.pdf)
+    - Word (.docx, .doc)
+    - PowerPoint (.pptx, .ppt)
+    - Excel (.xlsx, .xls)
+    - Images (.jpg, .jpeg, .png) - with OCR
+    - HTML (.html, .htm)
+    - Text (.txt, .md)
+    - Audio (.mp3, .wav) - with transcription
+    """
+
+    def __init__(self):
+        """Initialize markitdown parser."""
+        # Lazy import to avoid requiring markitdown for all users
+        try:
+            from markitdown import MarkItDown
+
+            self._markitdown = MarkItDown()
+        except ImportError as e:
+            raise ImportError(
+                "markitdown package is required for file parsing. Install with: pip install markitdown"
+            ) from e
+
+    async def convert(self, file_data: bytes, filename: str) -> str:
+        """Parse file to markdown using markitdown."""
+        # markitdown is synchronous, so we run it in executor to avoid blocking
+        loop = asyncio.get_event_loop()
+        return await loop.run_in_executor(None, self._convert_sync, file_data, filename)
+
+    def _convert_sync(self, file_data: bytes, filename: str) -> str:
+        """Synchronous parsing (runs in thread pool)."""
+        # Write to temp file (markitdown requires file path)
+        with tempfile.NamedTemporaryFile(suffix=Path(filename).suffix, delete=False) as tmp:
+            tmp.write(file_data)
+            tmp_path = tmp.name
+
+        try:
+            # Parse using markitdown
+            result = self._markitdown.convert(tmp_path)
+
+            if not result or not result.text_content:
+                raise RuntimeError(f"No content extracted from '{filename}'")
+
+            return result.text_content
+
+        except Exception as e:
+            logger.error(f"Markitdown parsing failed for {filename}: {e}")
+            raise RuntimeError(f"Failed to parse '{filename}': {e}") from e
+
+        finally:
+            # Clean up temp file
+            try:
+                Path(tmp_path).unlink()
+            except Exception:
+                pass
+
+    def supports(self, filename: str, content_type: str | None = None) -> bool:
+        """Check if markitdown supports this file type."""
+        # Supported extensions (from markitdown docs)
+        supported_extensions = {
+            # Documents
+            ".pdf",
+            ".docx",
+            ".doc",
+            ".pptx",
+            ".ppt",
+            ".xlsx",
+            ".xls",
+            # Images (with OCR)
+            ".jpg",
+            ".jpeg",
+            ".png",
+            # Web
+            ".html",
+            ".htm",
+            # Text
+            ".txt",
+            ".md",
+            ".csv",
+            # Audio (with transcription)
+            ".mp3",
+            ".wav",
+        }
+
+        ext = Path(filename).suffix.lower()
+        return ext in supported_extensions
+
+    def name(self) -> str:
+        """Get parser name."""
+        return "markitdown"

hindsight-api/hindsight_api/engine/providers/__init__.py

@@ -0,0 +1,14 @@
+"""
+LLM provider implementations.
+
+This package contains concrete implementations of the LLMInterface for various providers.
+"""
+
+from .anthropic_llm import AnthropicLLM
+from .claude_code_llm import ClaudeCodeLLM
+from .codex_llm import CodexLLM
+from .gemini_llm import GeminiLLM
+from .mock_llm import MockLLM
+from .openai_compatible_llm import OpenAICompatibleLLM
+
+__all__ = ["AnthropicLLM", "ClaudeCodeLLM", "CodexLLM", "GeminiLLM", "MockLLM", "OpenAICompatibleLLM"]

hindsight-api/hindsight_api/engine/providers/anthropic_llm.py

@@ -0,0 +1,477 @@
+"""
+Anthropic LLM provider using the Anthropic Python SDK.
+
+This provider enables using Claude models from Anthropic with support for:
+- Structured JSON output
+- Tool/function calling with proper format conversion
+- Extended thinking mode
+- Retry logic with exponential backoff
+"""
+
+import asyncio
+import json
+import logging
+import time
+from typing import Any
+
+from hindsight_api.engine.llm_interface import LLMInterface, OutputTooLongError
+from hindsight_api.engine.response_models import LLMToolCall, LLMToolCallResult, TokenUsage
+from hindsight_api.metrics import get_metrics_collector
+
+logger = logging.getLogger(__name__)
+
+
+class AnthropicLLM(LLMInterface):
+    """
+    LLM provider using Anthropic's Claude models.
+
+    Supports structured output, tool calling, and extended thinking mode.
+    Handles format conversion between OpenAI-style messages and Anthropic's format.
+    """
+
+    def __init__(
+        self,
+        provider: str,
+        api_key: str,
+        base_url: str,
+        model: str,
+        reasoning_effort: str = "low",
+        timeout: float = 300.0,
+        **kwargs: Any,
+    ):
+        """
+        Initialize Anthropic LLM provider.
+
+        Args:
+            provider: Provider name (should be "anthropic").
+            api_key: Anthropic API key.
+            base_url: Base URL for the API (optional, uses Anthropic default if empty).
+            model: Model name (e.g., "claude-sonnet-4-20250514").
+            reasoning_effort: Reasoning effort level (not used by Anthropic).
+            timeout: Request timeout in seconds.
+            **kwargs: Additional provider-specific parameters.
+        """
+        super().__init__(provider, api_key, base_url, model, reasoning_effort, **kwargs)
+
+        if not self.api_key:
+            raise ValueError("API key is required for Anthropic provider")
+
+        # Import and initialize Anthropic client
+        try:
+            from anthropic import AsyncAnthropic
+
+            client_kwargs: dict[str, Any] = {"api_key": self.api_key}
+            if self.base_url:
+                client_kwargs["base_url"] = self.base_url
+            if timeout:
+                client_kwargs["timeout"] = timeout
+
+            self._client = AsyncAnthropic(**client_kwargs)
+            logger.info(f"Anthropic client initialized for model: {self.model}")
+        except ImportError as e:
+            raise RuntimeError("Anthropic SDK not installed. Run: uv add anthropic or pip install anthropic") from e
+
+    async def verify_connection(self) -> None:
+        """
+        Verify that the Anthropic provider is configured correctly by making a simple test call.
+
+        Raises:
+            RuntimeError: If the connection test fails.
+        """
+        try:
+            test_messages = [{"role": "user", "content": "test"}]
+            await self.call(
+                messages=test_messages,
+                max_completion_tokens=10,
+                temperature=0.0,
+                scope="verification",
+                max_retries=0,
+            )
+            logger.info("Anthropic connection verified successfully")
+        except Exception as e:
+            logger.error(f"Anthropic connection verification failed: {e}")
+            raise RuntimeError(f"Failed to verify Anthropic connection: {e}") from e
+
+    async def call(
+        self,
+        messages: list[dict[str, str]],
+        response_format: Any | None = None,
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "memory",
+        max_retries: int = 10,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 60.0,
+        skip_validation: bool = False,
+        strict_schema: bool = False,
+        return_usage: bool = False,
+    ) -> Any:
+        """
+        Make an LLM API call with retry logic.
+
+        Args:
+            messages: List of message dicts with 'role' and 'content'.
+            response_format: Optional Pydantic model for structured output.
+            max_completion_tokens: Maximum tokens in response.
+            temperature: Sampling temperature (0.0-2.0).
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            skip_validation: Return raw JSON without Pydantic validation.
+            strict_schema: Use strict JSON schema enforcement (not supported by Anthropic).
+            return_usage: If True, return tuple (result, TokenUsage) instead of just result.
+
+        Returns:
+            If return_usage=False: Parsed response if response_format is provided, otherwise text content.
+            If return_usage=True: Tuple of (result, TokenUsage) with token counts.
+
+        Raises:
+            OutputTooLongError: If output exceeds token limits.
+            Exception: Re-raises API errors after retries exhausted.
+        """
+        from anthropic import APIConnectionError, APIStatusError, RateLimitError
+
+        start_time = time.time()
+
+        # Convert OpenAI-style messages to Anthropic format
+        system_prompt = None
+        anthropic_messages = []
+
+        for msg in messages:
+            role = msg.get("role", "user")
+            content = msg.get("content", "")
+
+            if role == "system":
+                if system_prompt:
+                    system_prompt += "\n\n" + content
+                else:
+                    system_prompt = content
+            else:
+                anthropic_messages.append({"role": role, "content": content})
+
+        # Add JSON schema instruction if response_format is provided
+        if response_format is not None and hasattr(response_format, "model_json_schema"):
+            schema = response_format.model_json_schema()
+            schema_msg = f"\n\nYou must respond with valid JSON matching this schema:\n{json.dumps(schema, indent=2)}"
+            if system_prompt:
+                system_prompt += schema_msg
+            else:
+                system_prompt = schema_msg
+
+        # Prepare parameters
+        call_params: dict[str, Any] = {
+            "model": self.model,
+            "messages": anthropic_messages,
+            "max_tokens": max_completion_tokens if max_completion_tokens is not None else 4096,
+        }
+
+        if system_prompt:
+            call_params["system"] = system_prompt
+
+        if temperature is not None:
+            call_params["temperature"] = temperature
+
+        last_exception = None
+
+        for attempt in range(max_retries + 1):
+            try:
+                response = await self._client.messages.create(**call_params)
+
+                # Anthropic response content is a list of blocks
+                content = ""
+                for block in response.content:
+                    if block.type == "text":
+                        content += block.text
+
+                if response_format is not None:
+                    # Models may wrap JSON in markdown code blocks
+                    clean_content = content
+                    if "```json" in content:
+                        clean_content = content.split("```json")[1].split("```")[0].strip()
+                    elif "```" in content:
+                        clean_content = content.split("```")[1].split("```")[0].strip()
+
+                    try:
+                        json_data = json.loads(clean_content)
+                    except json.JSONDecodeError:
+                        # Fallback to parsing raw content if markdown stripping failed
+                        json_data = json.loads(content)
+
+                    if skip_validation:
+                        result = json_data
+                    else:
+                        result = response_format.model_validate(json_data)
+                else:
+                    result = content
+
+                # Record metrics and log slow calls
+                duration = time.time() - start_time
+                input_tokens = response.usage.input_tokens or 0 if response.usage else 0
+                output_tokens = response.usage.output_tokens or 0 if response.usage else 0
+                total_tokens = input_tokens + output_tokens
+
+                # Record LLM metrics
+                metrics = get_metrics_collector()
+                metrics.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    duration=duration,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    success=True,
+                )
+
+                # Record trace span
+                from hindsight_api.tracing import _serialize_for_span, get_span_recorder
+
+                finish_reason = response.stop_reason if hasattr(response, "stop_reason") else None
+                span_recorder = get_span_recorder()
+                span_recorder.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    messages=messages,
+                    response_content=_serialize_for_span(result),
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    duration=duration,
+                    finish_reason=finish_reason,
+                    error=None,
+                )
+
+                # Log slow calls
+                if duration > 10.0:
+                    logger.info(
+                        f"slow llm call: scope={scope}, model={self.provider}/{self.model}, "
+                        f"input_tokens={input_tokens}, output_tokens={output_tokens}, "
+                        f"time={duration:.3f}s"
+                    )
+
+                if return_usage:
+                    token_usage = TokenUsage(
+                        input_tokens=input_tokens,
+                        output_tokens=output_tokens,
+                        total_tokens=total_tokens,
+                    )
+                    return result, token_usage
+                return result
+
+            except json.JSONDecodeError as e:
+                last_exception = e
+                if attempt < max_retries:
+                    logger.warning("Anthropic returned invalid JSON, retrying...")
+                    backoff = min(initial_backoff * (2**attempt), max_backoff)
+                    await asyncio.sleep(backoff)
+                    continue
+                else:
+                    logger.error(f"Anthropic returned invalid JSON after {max_retries + 1} attempts")
+                    raise
+
+            except (APIConnectionError, RateLimitError, APIStatusError) as e:
+                # Fast fail on 401/403
+                if isinstance(e, APIStatusError) and e.status_code in (401, 403):
+                    logger.error(f"Anthropic auth error (HTTP {e.status_code}), not retrying: {str(e)}")
+                    raise
+
+                last_exception = e
+                if attempt < max_retries:
+                    # Check if it's a rate limit or server error
+                    should_retry = isinstance(e, (APIConnectionError, RateLimitError)) or (
+                        isinstance(e, APIStatusError) and e.status_code >= 500
+                    )
+
+                    if should_retry:
+                        backoff = min(initial_backoff * (2**attempt), max_backoff)
+                        jitter = backoff * 0.2 * (2 * (time.time() % 1) - 1)
+                        await asyncio.sleep(backoff + jitter)
+                        continue
+
+                logger.error(f"Anthropic API error after {max_retries + 1} attempts: {str(e)}")
+                raise
+
+            except Exception as e:
+                logger.error(f"Unexpected error during Anthropic call: {type(e).__name__}: {str(e)}")
+                raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("Anthropic call failed after all retries")
+
+    async def call_with_tools(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "tools",
+        max_retries: int = 5,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 30.0,
+        tool_choice: str | dict[str, Any] = "auto",
+    ) -> LLMToolCallResult:
+        """
+        Make an LLM API call with tool/function calling support.
+
+        Args:
+            messages: List of message dicts. Can include tool results with role='tool'.
+            tools: List of tool definitions in OpenAI format.
+            max_completion_tokens: Maximum tokens in response.
+            temperature: Sampling temperature (0.0-2.0).
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            tool_choice: How to choose tools - "auto", "none", "required", or specific function.
+
+        Returns:
+            LLMToolCallResult with content and/or tool_calls.
+        """
+        from anthropic import APIConnectionError, APIStatusError
+
+        start_time = time.time()
+
+        # Convert OpenAI tool format to Anthropic format
+        anthropic_tools = []
+        for tool in tools:
+            func = tool.get("function", {})
+            anthropic_tools.append(
+                {
+                    "name": func.get("name", ""),
+                    "description": func.get("description", ""),
+                    "input_schema": func.get("parameters", {"type": "object", "properties": {}}),
+                }
+            )
+
+        # Convert messages - handle tool results
+        system_prompt = None
+        anthropic_messages = []
+        for msg in messages:
+            role = msg.get("role", "user")
+            content = msg.get("content", "")
+
+            if role == "system":
+                system_prompt = (system_prompt + "\n\n" + content) if system_prompt else content
+            elif role == "tool":
+                # Anthropic uses tool_result blocks
+                anthropic_messages.append(
+                    {
+                        "role": "user",
+                        "content": [
+                            {"type": "tool_result", "tool_use_id": msg.get("tool_call_id", ""), "content": content}
+                        ],
+                    }
+                )
+            elif role == "assistant" and msg.get("tool_calls"):
+                # Convert assistant tool calls
+                tool_use_blocks = []
+                for tc in msg["tool_calls"]:
+                    tool_use_blocks.append(
+                        {
+                            "type": "tool_use",
+                            "id": tc.get("id", ""),
+                            "name": tc.get("function", {}).get("name", ""),
+                            "input": json.loads(tc.get("function", {}).get("arguments", "{}")),
+                        }
+                    )
+                anthropic_messages.append({"role": "assistant", "content": tool_use_blocks})
+            else:
+                anthropic_messages.append({"role": role, "content": content})
+
+        call_params: dict[str, Any] = {
+            "model": self.model,
+            "messages": anthropic_messages,
+            "tools": anthropic_tools,
+            "max_tokens": max_completion_tokens or 4096,
+        }
+        if system_prompt:
+            call_params["system"] = system_prompt
+
+        if temperature is not None:
+            call_params["temperature"] = temperature
+
+        last_exception = None
+        for attempt in range(max_retries + 1):
+            try:
+                response = await self._client.messages.create(**call_params)
+
+                # Extract content and tool calls
+                content_parts = []
+                tool_calls: list[LLMToolCall] = []
+
+                for block in response.content:
+                    if block.type == "text":
+                        content_parts.append(block.text)
+                    elif block.type == "tool_use":
+                        tool_calls.append(LLMToolCall(id=block.id, name=block.name, arguments=block.input or {}))
+
+                content = "".join(content_parts) if content_parts else None
+                finish_reason = "tool_calls" if tool_calls else "stop"
+
+                # Extract token usage
+                input_tokens = response.usage.input_tokens or 0
+                output_tokens = response.usage.output_tokens or 0
+
+                # Record metrics
+                metrics = get_metrics_collector()
+                duration = time.time() - start_time
+                metrics.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    duration=duration,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    success=True,
+                )
+
+                # Record OpenTelemetry span
+                from hindsight_api.tracing import get_span_recorder
+
+                span_recorder = get_span_recorder()
+                # Convert LLMToolCall objects to dicts for span recording
+                tool_calls_dict = (
+                    [{"id": tc.id, "name": tc.name, "arguments": tc.arguments} for tc in tool_calls]
+                    if tool_calls
+                    else None
+                )
+                span_recorder.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    messages=messages,
+                    response_content=content,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    duration=duration,
+                    finish_reason=finish_reason,
+                    error=None,
+                    tool_calls=tool_calls_dict,
+                )
+
+                return LLMToolCallResult(
+                    content=content,
+                    tool_calls=tool_calls,
+                    finish_reason=finish_reason,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                )
+
+            except (APIConnectionError, APIStatusError) as e:
+                if isinstance(e, APIStatusError) and e.status_code in (401, 403):
+                    raise
+                last_exception = e
+                if attempt < max_retries:
+                    await asyncio.sleep(min(initial_backoff * (2**attempt), max_backoff))
+                    continue
+                raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("Anthropic tool call failed")
+
+    async def cleanup(self) -> None:
+        """Clean up resources (close Anthropic client connections)."""
+        if hasattr(self, "_client") and self._client:
+            await self._client.close()

hindsight-api/hindsight_api/engine/providers/claude_code_llm.py

@@ -0,0 +1,510 @@
+"""
+Claude Code LLM provider using Claude Agent SDK.
+
+This provider enables using Claude Pro/Max subscriptions for API calls
+via the Claude CLI authentication. It uses the Claude Agent SDK which
+automatically handles authentication via `claude auth login` credentials.
+"""
+
+import asyncio
+import json
+import logging
+import time
+from typing import Any
+
+from hindsight_api.engine.llm_interface import LLMInterface, OutputTooLongError
+from hindsight_api.engine.response_models import LLMToolCall, LLMToolCallResult, TokenUsage
+from hindsight_api.metrics import get_metrics_collector
+
+logger = logging.getLogger(__name__)
+
+
+class ClaudeCodeLLM(LLMInterface):
+    """
+    LLM provider using Claude Code authentication.
+
+    Authenticates using Claude Pro/Max credentials via `claude auth login`
+    and makes API calls through the Claude Agent SDK.
+    """
+
+    def __init__(
+        self,
+        provider: str,
+        api_key: str,  # Will be ignored, uses CLI auth
+        base_url: str,
+        model: str,
+        reasoning_effort: str = "low",
+        **kwargs: Any,
+    ):
+        """Initialize Claude Code LLM provider."""
+        super().__init__(provider, api_key, base_url, model, reasoning_effort, **kwargs)
+
+        # Verify Claude Agent SDK is available
+        try:
+            self._verify_claude_code_available()
+            logger.info("Claude Code: Using Claude Agent SDK (authentication via claude auth login)")
+        except Exception as e:
+            raise RuntimeError(
+                f"Failed to initialize Claude Code provider: {e}\n\n"
+                "To set up Claude Code authentication:\n"
+                "1. Install Claude Code CLI: npm install -g @anthropics/claude-code\n"
+                "2. Login with your Pro/Max plan: claude auth login\n"
+                "3. Verify authentication: claude --version\n\n"
+                "Or use a different provider (anthropic, openai, gemini) with API keys."
+            ) from e
+
+        # Metrics collector is imported at module level
+
+    def _verify_claude_code_available(self) -> None:
+        """
+        Verify that Claude Agent SDK can be imported and is properly configured.
+
+        Raises:
+            ImportError: If Claude Agent SDK is not installed.
+            RuntimeError: If Claude Code is not authenticated.
+        """
+        try:
+            # Import Claude Agent SDK
+            # Reduce Claude Agent SDK logging verbosity
+            import logging as sdk_logging
+
+            from claude_agent_sdk import query  # noqa: F401
+
+            sdk_logging.getLogger("claude_agent_sdk").setLevel(sdk_logging.WARNING)
+            sdk_logging.getLogger("claude_agent_sdk._internal").setLevel(sdk_logging.WARNING)
+
+            logger.debug("Claude Agent SDK imported successfully")
+        except ImportError as e:
+            raise ImportError(
+                "Claude Agent SDK not installed. Run: uv add claude-agent-sdk or pip install claude-agent-sdk"
+            ) from e
+
+        # SDK will automatically check for authentication when first used
+        # No need to verify here - let it fail gracefully on first call with helpful error
+
+    async def verify_connection(self) -> None:
+        """
+        Verify that the Claude Code provider is configured correctly by making a simple test call.
+
+        Raises:
+            RuntimeError: If the connection test fails.
+        """
+        try:
+            test_messages = [{"role": "user", "content": "test"}]
+            await self.call(
+                messages=test_messages,
+                max_completion_tokens=10,
+                temperature=0.0,
+                scope="verification",
+                max_retries=0,
+            )
+            logger.info("Claude Code connection verified successfully")
+        except Exception as e:
+            logger.error(f"Claude Code connection verification failed: {e}")
+            raise RuntimeError(f"Failed to verify Claude Code connection: {e}") from e
+
+    async def call(
+        self,
+        messages: list[dict[str, str]],
+        response_format: Any | None = None,
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "memory",
+        max_retries: int = 10,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 60.0,
+        skip_validation: bool = False,
+        strict_schema: bool = False,
+        return_usage: bool = False,
+    ) -> Any:
+        """
+        Make an LLM API call with retry logic.
+
+        Args:
+            messages: List of message dicts with 'role' and 'content'.
+            response_format: Optional Pydantic model for structured output.
+            max_completion_tokens: Maximum tokens in response (ignored by Claude Agent SDK).
+            temperature: Sampling temperature (ignored by Claude Agent SDK).
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            skip_validation: Return raw JSON without Pydantic validation.
+            strict_schema: Use strict JSON schema enforcement (not supported).
+            return_usage: If True, return tuple (result, TokenUsage) instead of just result.
+
+        Returns:
+            If return_usage=False: Parsed response if response_format is provided, otherwise text content.
+            If return_usage=True: Tuple of (result, TokenUsage) with estimated token counts.
+
+        Raises:
+            OutputTooLongError: If output exceeds token limits (not supported by Claude Agent SDK).
+            Exception: Re-raises API errors after retries exhausted.
+        """
+        from claude_agent_sdk import AssistantMessage, ClaudeAgentOptions, TextBlock, query
+
+        start_time = time.time()
+
+        # Build system prompt
+        system_prompt = ""
+        user_content = ""
+
+        for msg in messages:
+            role = msg.get("role", "user")
+            content = msg.get("content", "")
+
+            if role == "system":
+                system_prompt += ("\n\n" + content) if system_prompt else content
+            elif role == "user":
+                user_content += ("\n\n" + content) if user_content else content
+            elif role == "assistant":
+                # Claude Agent SDK doesn't support multi-turn easily in query()
+                # For now, prepend assistant messages to user content
+                user_content += f"\n\n[Previous assistant response: {content}]"
+
+        # Add JSON schema instruction if response_format is provided
+        if response_format is not None and hasattr(response_format, "model_json_schema"):
+            schema = response_format.model_json_schema()
+            schema_instruction = (
+                f"\n\nYou must respond with valid JSON matching this schema:\n{json.dumps(schema, indent=2)}\n\n"
+                "Respond with ONLY the JSON, no markdown formatting."
+            )
+            user_content += schema_instruction
+
+        # Configure SDK options
+        options = ClaudeAgentOptions(
+            system_prompt=system_prompt if system_prompt else None,
+            max_turns=1,  # Single-turn for API-style interactions
+            allowed_tools=[],  # Disable tools for standard LLM calls
+        )
+
+        # Call Claude Agent SDK
+        last_exception = None
+        for attempt in range(max_retries + 1):
+            try:
+                # Collect streaming response
+                full_text = ""
+
+                async for message in query(prompt=user_content, options=options):
+                    if isinstance(message, AssistantMessage):
+                        for block in message.content:
+                            if isinstance(block, TextBlock):
+                                full_text += block.text
+
+                # Handle structured output
+                if response_format is not None:
+                    # Models may wrap JSON in markdown
+                    clean_text = full_text
+                    if "```json" in full_text:
+                        clean_text = full_text.split("```json")[1].split("```")[0].strip()
+                    elif "```" in full_text:
+                        clean_text = full_text.split("```")[1].split("```")[0].strip()
+
+                    try:
+                        json_data = json.loads(clean_text)
+                    except json.JSONDecodeError as e:
+                        logger.warning(f"Claude Code JSON parse error (attempt {attempt + 1}/{max_retries + 1}): {e}")
+                        if attempt < max_retries:
+                            backoff = min(initial_backoff * (2**attempt), max_backoff)
+                            await asyncio.sleep(backoff)
+                            last_exception = e
+                            continue
+                        raise
+
+                    if skip_validation:
+                        result = json_data
+                    else:
+                        result = response_format.model_validate(json_data)
+                else:
+                    result = full_text
+
+                # Record metrics
+                duration = time.time() - start_time
+                metrics = get_metrics_collector()
+
+                # Estimate token usage (Claude Agent SDK doesn't report exact counts)
+                # Use character count / 4 as rough estimate (1 token ≈ 4 characters)
+                estimated_input = sum(len(m.get("content", "")) for m in messages) // 4
+                estimated_output = len(full_text) // 4
+
+                metrics.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    duration=duration,
+                    input_tokens=estimated_input,
+                    output_tokens=estimated_output,
+                    success=True,
+                )
+
+                # Record trace span
+                from hindsight_api.tracing import get_span_recorder
+
+                span_recorder = get_span_recorder()
+                span_recorder.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    messages=messages,
+                    response_content=result if isinstance(result, str) else json.dumps(result),
+                    input_tokens=estimated_input,
+                    output_tokens=estimated_output,
+                    duration=duration,
+                    finish_reason=None,
+                    error=None,
+                )
+
+                # Log slow calls
+                if duration > 10.0:
+                    logger.info(
+                        f"slow llm call: scope={scope}, model={self.provider}/{self.model}, time={duration:.3f}s"
+                    )
+
+                if return_usage:
+                    token_usage = TokenUsage(
+                        input_tokens=estimated_input,
+                        output_tokens=estimated_output,
+                        total_tokens=estimated_input + estimated_output,
+                    )
+                    return result, token_usage
+
+                return result
+
+            except Exception as e:
+                last_exception = e
+
+                # Check for authentication errors
+                error_str = str(e).lower()
+                if "auth" in error_str or "login" in error_str or "credential" in error_str:
+                    logger.error(f"Claude Code authentication error: {e}")
+                    raise RuntimeError(
+                        f"Claude Code authentication failed: {e}\n\n"
+                        "Run 'claude auth login' to authenticate with Claude Pro/Max."
+                    ) from e
+
+                if attempt < max_retries:
+                    backoff = min(initial_backoff * (2**attempt), max_backoff)
+                    logger.warning(f"Claude Code error (attempt {attempt + 1}/{max_retries + 1}): {e}")
+                    await asyncio.sleep(backoff)
+                    continue
+                else:
+                    logger.error(f"Claude Code error after {max_retries + 1} attempts: {e}")
+                    raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("Claude Code call failed after all retries")
+
+    async def call_with_tools(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "tools",
+        max_retries: int = 5,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 30.0,
+        tool_choice: str | dict[str, Any] = "auto",
+    ) -> LLMToolCallResult:
+        """
+        Make an LLM API call with tool/function calling support using Claude Agent SDK.
+
+        This implementation uses ClaudeSDKClient (not query()) because custom tools via
+        SDK MCP servers are only supported with the client. Tools are converted from OpenAI
+        format to SDK MCP tools, and tool names are formatted as mcp__hindsight_tools__{name}.
+
+        Args:
+            messages: List of message dicts. Can include tool results with role='tool'.
+            tools: List of tool definitions in OpenAI format.
+            max_completion_tokens: Maximum tokens in response (not used by Claude Agent SDK).
+            temperature: Sampling temperature (not used by Claude Agent SDK).
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            tool_choice: How to choose tools (not used by Claude Agent SDK).
+
+        Returns:
+            LLMToolCallResult with content and/or tool_calls.
+        """
+        from claude_agent_sdk import (
+            AssistantMessage,
+            ClaudeAgentOptions,
+            ClaudeSDKClient,
+            SdkMcpTool,
+            TextBlock,
+            ToolUseBlock,
+            create_sdk_mcp_server,
+        )
+
+        start_time = time.time()
+
+        # Convert OpenAI tool format to Claude Agent SDK SdkMcpTool format
+        sdk_tools: list[SdkMcpTool] = []
+        tool_names: list[str] = []
+
+        for tool in tools:
+            func = tool.get("function", {})
+            tool_name = func.get("name", "")
+            tool_description = func.get("description", "")
+            parameters = func.get("parameters", {})
+
+            # Create a handler with proper closure to avoid transport issues
+            def make_handler(name: str):
+                async def handler(args: dict[str, Any]) -> dict[str, Any]:
+                    # Return immediately with success - tool execution happens externally
+                    return {
+                        "content": [
+                            {
+                                "type": "text",
+                                "text": f"[Tool {name} called successfully]",
+                            }
+                        ]
+                    }
+
+                return handler
+
+            sdk_tools.append(
+                SdkMcpTool(
+                    name=tool_name,
+                    description=tool_description,
+                    input_schema=parameters,
+                    handler=make_handler(tool_name),
+                )
+            )
+            tool_names.append(tool_name)
+
+        # Create an MCP server with the tools
+        mcp_server = create_sdk_mcp_server(
+            name="hindsight_tools",
+            version="1.0.0",
+            tools=sdk_tools if sdk_tools else None,
+        )
+
+        # Build system prompt and user content from messages
+        system_prompt = ""
+        user_content = ""
+
+        for msg in messages:
+            role = msg.get("role", "user")
+            content = msg.get("content", "")
+
+            if role == "system":
+                system_prompt += ("\n\n" + content) if system_prompt else content
+            elif role == "user":
+                user_content += ("\n\n" + content) if user_content else content
+            elif role == "assistant":
+                # Include previous assistant messages as context
+                user_content += f"\n\n[Previous assistant response: {content}]"
+            elif role == "tool":
+                # Tool results are already in tool_results_map, append to user context
+                tool_call_id = msg.get("tool_call_id", "")
+                user_content += f"\n\n[Tool result for {tool_call_id}: {content}]"
+
+        # Format tool names for SDK MCP servers: mcp__{server_name}__{tool_name}
+        # This is required by the Claude Agent SDK for MCP server tools
+        allowed_tool_names = [f"mcp__hindsight_tools__{name}" for name in tool_names]
+
+        # Configure SDK options with MCP server
+        options = ClaudeAgentOptions(
+            system_prompt=system_prompt if system_prompt else None,
+            max_turns=1,  # Single-turn for API-style interactions
+            mcp_servers={"hindsight_tools": mcp_server} if sdk_tools else {},
+            allowed_tools=allowed_tool_names if allowed_tool_names else [],
+        )
+
+        # Call Claude Agent SDK with retry logic
+        last_exception = None
+        for attempt in range(max_retries + 1):
+            try:
+                full_text = ""
+                tool_calls: list[LLMToolCall] = []
+
+                # Use ClaudeSDKClient for tool calling support
+                # Note: query() does NOT support custom tools, only ClaudeSDKClient does
+                async with ClaudeSDKClient(options=options) as client:
+                    # Send the query
+                    await client.query(user_content)
+
+                    # Receive response
+                    async for message in client.receive_response():
+                        if isinstance(message, AssistantMessage):
+                            for block in message.content:
+                                if isinstance(block, TextBlock):
+                                    full_text += block.text
+                                elif isinstance(block, ToolUseBlock):
+                                    # SDK returns tool names with MCP prefix (mcp__hindsight_tools__{name})
+                                    # Strip the prefix to return original tool name expected by caller
+                                    tool_name = block.name
+                                    if tool_name.startswith("mcp__hindsight_tools__"):
+                                        tool_name = tool_name.replace("mcp__hindsight_tools__", "", 1)
+
+                                    tool_calls.append(
+                                        LLMToolCall(
+                                            id=block.id,
+                                            name=tool_name,
+                                            arguments=block.input,
+                                        )
+                                    )
+
+                # Record metrics
+                duration = time.time() - start_time
+                metrics = get_metrics_collector()
+
+                # Estimate token usage (Claude Agent SDK doesn't report exact counts)
+                estimated_input = sum(len(m.get("content", "")) for m in messages) // 4
+                estimated_output = len(full_text) // 4
+
+                metrics.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    duration=duration,
+                    input_tokens=estimated_input,
+                    output_tokens=estimated_output,
+                    success=True,
+                )
+
+                # Log slow calls
+                if duration > 10.0:
+                    logger.info(
+                        f"slow llm call: scope={scope}, model={self.provider}/{self.model}, time={duration:.3f}s"
+                    )
+
+                return LLMToolCallResult(
+                    content=full_text if full_text else None,
+                    tool_calls=tool_calls,
+                    finish_reason="tool_calls" if tool_calls else "stop",
+                    input_tokens=estimated_input,
+                    output_tokens=estimated_output,
+                )
+
+            except Exception as e:
+                last_exception = e
+
+                # Check for authentication errors
+                error_str = str(e).lower()
+                if "auth" in error_str or "login" in error_str or "credential" in error_str:
+                    logger.error(f"Claude Code authentication error: {e}")
+                    raise RuntimeError(
+                        f"Claude Code authentication failed: {e}\n\n"
+                        "Run 'claude auth login' to authenticate with Claude Pro/Max."
+                    ) from e
+
+                if attempt < max_retries:
+                    backoff = min(initial_backoff * (2**attempt), max_backoff)
+                    logger.warning(f"Claude Code tool call error (attempt {attempt + 1}/{max_retries + 1}): {e}")
+                    await asyncio.sleep(backoff)
+                    continue
+                else:
+                    logger.error(f"Claude Code tool call error after {max_retries + 1} attempts: {e}")
+                    raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("Claude Code tool call failed after all retries")
+
+    async def cleanup(self) -> None:
+        """Clean up resources (no HTTP client to close for Claude Agent SDK)."""
+        pass

hindsight-api/hindsight_api/engine/providers/codex_llm.py

@@ -0,0 +1,621 @@
+"""
+OpenAI Codex LLM provider using ChatGPT Plus/Pro OAuth authentication.
+
+This provider enables using ChatGPT Plus/Pro subscriptions for API calls
+without separate OpenAI Platform API credits. It uses OAuth tokens from
+~/.codex/auth.json and communicates with the ChatGPT backend API.
+"""
+
+import asyncio
+import json
+import logging
+import os
+import time
+import uuid
+from pathlib import Path
+from typing import Any
+
+import httpx
+
+from hindsight_api.engine.llm_interface import LLMInterface, OutputTooLongError
+from hindsight_api.engine.response_models import LLMToolCall, LLMToolCallResult, TokenUsage
+from hindsight_api.metrics import get_metrics_collector
+
+logger = logging.getLogger(__name__)
+
+
+class CodexLLM(LLMInterface):
+    """
+    LLM provider using OpenAI Codex OAuth authentication.
+
+    Authenticates using ChatGPT Plus/Pro credentials stored in ~/.codex/auth.json
+    and makes API calls to chatgpt.com/backend-api/codex/responses.
+    """
+
+    def __init__(
+        self,
+        provider: str,
+        api_key: str,  # Will be ignored, reads from ~/.codex/auth.json
+        base_url: str,
+        model: str,
+        reasoning_effort: str = "low",
+        **kwargs: Any,
+    ):
+        """Initialize Codex LLM provider."""
+        super().__init__(provider, api_key, base_url, model, reasoning_effort, **kwargs)
+
+        # Load Codex OAuth credentials
+        try:
+            self.access_token, self.account_id = self._load_codex_auth()
+            logger.info(f"Loaded Codex OAuth credentials for account: {self.account_id}")
+        except Exception as e:
+            raise RuntimeError(
+                f"Failed to load Codex OAuth credentials from ~/.codex/auth.json: {e}\n\n"
+                "To set up Codex authentication:\n"
+                "1. Install Codex CLI: npm install -g @openai/codex\n"
+                "2. Login: codex auth login\n"
+                "3. Verify: ls ~/.codex/auth.json\n\n"
+                "Or use a different provider (openai, anthropic, gemini) with API keys."
+            ) from e
+
+        # Use ChatGPT backend API endpoint
+        if not self.base_url:
+            self.base_url = "https://chatgpt.com/backend-api"
+
+        # Normalize model name (strip openai/ prefix if present)
+        if self.model.startswith("openai/"):
+            self.model = self.model[len("openai/") :]
+
+        # Map reasoning effort to Codex reasoning summary format
+        # Codex supports: "auto", "concise", "detailed"
+        self.reasoning_summary = self._map_reasoning_effort(reasoning_effort)
+
+        # HTTP client for SSE streaming
+        self._client = httpx.AsyncClient(timeout=120.0)
+
+    def _load_codex_auth(self) -> tuple[str, str]:
+        """
+        Load OAuth credentials from ~/.codex/auth.json.
+
+        Returns:
+            Tuple of (access_token, account_id).
+
+        Raises:
+            FileNotFoundError: If auth file doesn't exist.
+            ValueError: If auth file is invalid.
+        """
+        auth_file = Path.home() / ".codex" / "auth.json"
+
+        if not auth_file.exists():
+            raise FileNotFoundError(
+                f"Codex auth file not found: {auth_file}\nRun 'codex auth login' to authenticate with ChatGPT Plus/Pro."
+            )
+
+        with open(auth_file) as f:
+            data = json.load(f)
+
+        # Validate auth structure
+        auth_mode = data.get("auth_mode")
+        if auth_mode != "chatgpt":
+            raise ValueError(f"Expected auth_mode='chatgpt', got: {auth_mode}")
+
+        tokens = data.get("tokens", {})
+        access_token = tokens.get("access_token")
+        account_id = tokens.get("account_id")
+
+        if not access_token:
+            raise ValueError("No access_token found in Codex auth file. Run 'codex auth login' again.")
+
+        return access_token, account_id
+
+    def _map_reasoning_effort(self, effort: str) -> str:
+        """
+        Map standard reasoning effort to Codex reasoning summary format.
+
+        Args:
+            effort: Standard effort level ("low", "medium", "high", "xhigh").
+
+        Returns:
+            Codex reasoning summary: "concise", "detailed", or "auto".
+        """
+        mapping = {
+            "low": "concise",
+            "medium": "auto",
+            "high": "detailed",
+            "xhigh": "detailed",
+        }
+        return mapping.get(effort.lower(), "auto")
+
+    async def verify_connection(self) -> None:
+        """Verify Codex connection by making a simple test call."""
+        try:
+            logger.info(f"Verifying Codex LLM: model={self.model}, account={self.account_id}...")
+            await self.call(
+                messages=[{"role": "user", "content": "Say 'ok'"}],
+                max_completion_tokens=10,
+                max_retries=2,
+                initial_backoff=0.5,
+                max_backoff=2.0,
+                scope="verification",
+            )
+            logger.info(f"Codex LLM verified: {self.model}")
+        except Exception as e:
+            raise RuntimeError(f"Codex LLM connection verification failed for {self.model}: {e}") from e
+
+    async def call(
+        self,
+        messages: list[dict[str, str]],
+        response_format: Any | None = None,
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "memory",
+        max_retries: int = 10,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 60.0,
+        skip_validation: bool = False,
+        strict_schema: bool = False,
+        return_usage: bool = False,
+    ) -> Any:
+        """Make API call to Codex backend with SSE streaming."""
+        start_time = time.time()
+
+        # Prepare system instructions
+        system_instruction = ""
+        user_messages = []
+
+        for msg in messages:
+            role = msg.get("role", "user")
+            content = msg.get("content", "")
+
+            if role == "system":
+                system_instruction += ("\n\n" + content) if system_instruction else content
+            else:
+                user_messages.append(msg)
+
+        # Add JSON schema instruction if response_format is provided
+        if response_format is not None and hasattr(response_format, "model_json_schema"):
+            schema = response_format.model_json_schema()
+            schema_msg = f"\n\nYou must respond with valid JSON matching this schema:\n{json.dumps(schema, indent=2)}"
+            system_instruction += schema_msg
+
+        # gpt-5.2-codex only supports "detailed" reasoning summary
+        reasoning_summary = "detailed" if "5.2" in self.model else self.reasoning_summary
+
+        # Build Codex request payload
+        payload = {
+            "model": self.model,
+            "instructions": system_instruction,
+            "input": [
+                {
+                    "type": "message",
+                    "role": msg.get("role", "user"),
+                    "content": msg.get("content", ""),
+                }
+                for msg in user_messages
+            ],
+            "tools": [],
+            "tool_choice": "auto",
+            "parallel_tool_calls": True,
+            "reasoning": {"summary": reasoning_summary},
+            "store": False,  # Codex uses stateless mode
+            "stream": True,  # SSE streaming
+            "include": ["reasoning.encrypted_content"],
+            "prompt_cache_key": str(uuid.uuid4()),
+        }
+
+        headers = {
+            "Authorization": f"Bearer {self.access_token}",
+            "Content-Type": "application/json",
+            "OpenAI-Account-ID": self.account_id,
+            "User-Agent": "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7)",
+            "Origin": "https://chatgpt.com",
+        }
+
+        url = f"{self.base_url}/codex/responses"
+        last_exception = None
+
+        for attempt in range(max_retries + 1):
+            try:
+                response = await self._client.post(url, json=payload, headers=headers, timeout=120.0)
+                response.raise_for_status()
+
+                # Parse SSE stream
+                content = await self._parse_sse_stream(response)
+
+                # Handle structured output
+                if response_format is not None:
+                    # Models may wrap JSON in markdown
+                    clean_content = content
+                    if "```json" in content:
+                        clean_content = content.split("```json")[1].split("```")[0].strip()
+                    elif "```" in content:
+                        clean_content = content.split("```")[1].split("```")[0].strip()
+
+                    try:
+                        json_data = json.loads(clean_content)
+                    except json.JSONDecodeError as e:
+                        logger.warning(f"Codex JSON parse error (attempt {attempt + 1}/{max_retries + 1}): {e}")
+                        if attempt < max_retries:
+                            backoff = min(initial_backoff * (2**attempt), max_backoff)
+                            await asyncio.sleep(backoff)
+                            last_exception = e
+                            continue
+                        raise
+
+                    if skip_validation:
+                        result = json_data
+                    else:
+                        result = response_format.model_validate(json_data)
+                else:
+                    result = content
+
+                # Record metrics
+                duration = time.time() - start_time
+                metrics = get_metrics_collector()
+                metrics.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    duration=duration,
+                    input_tokens=0,  # Codex doesn't report token counts in SSE
+                    output_tokens=0,
+                    success=True,
+                )
+
+                # Record trace span
+                from hindsight_api.tracing import get_span_recorder
+
+                # Estimate tokens for tracing
+                estimated_input = sum(len(m.get("content", "")) for m in messages) // 4
+                estimated_output = len(content) // 4
+                span_recorder = get_span_recorder()
+                span_recorder.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    messages=messages,
+                    response_content=result if isinstance(result, str) else json.dumps(result),
+                    input_tokens=estimated_input,
+                    output_tokens=estimated_output,
+                    duration=duration,
+                    finish_reason=None,
+                    error=None,
+                )
+
+                if return_usage:
+                    # Codex doesn't provide token counts, estimate based on content
+                    estimated_input = sum(len(m.get("content", "")) for m in messages) // 4
+                    estimated_output = len(content) // 4
+                    token_usage = TokenUsage(
+                        input_tokens=estimated_input,
+                        output_tokens=estimated_output,
+                        total_tokens=estimated_input + estimated_output,
+                    )
+                    return result, token_usage
+
+                return result
+
+            except httpx.HTTPStatusError as e:
+                last_exception = e
+                status_code = e.response.status_code
+
+                # Fast fail on auth errors
+                if status_code in (401, 403):
+                    logger.error(f"Codex auth error (HTTP {status_code}): {e.response.text[:200]}")
+                    raise RuntimeError(
+                        "Codex authentication failed. Your OAuth token may have expired.\n"
+                        "Run 'codex auth login' to re-authenticate."
+                    ) from e
+
+                # Log the actual error message from the API
+                error_detail = e.response.text[:500] if hasattr(e.response, "text") else str(e)
+
+                if attempt < max_retries:
+                    backoff = min(initial_backoff * (2**attempt), max_backoff)
+                    logger.warning(
+                        f"Codex HTTP error {status_code} (attempt {attempt + 1}/{max_retries + 1}): {error_detail}"
+                    )
+                    await asyncio.sleep(backoff)
+                    continue
+                else:
+                    logger.error(
+                        f"Codex HTTP error after {max_retries + 1} attempts: Status {status_code}, Detail: {error_detail}"
+                    )
+                    raise
+
+            except httpx.RequestError as e:
+                last_exception = e
+                if attempt < max_retries:
+                    backoff = min(initial_backoff * (2**attempt), max_backoff)
+                    logger.warning(f"Codex connection error (attempt {attempt + 1}/{max_retries + 1}): {e}")
+                    await asyncio.sleep(backoff)
+                    continue
+                else:
+                    logger.error(f"Codex connection error after {max_retries + 1} attempts: {e}")
+                    raise
+
+            except Exception as e:
+                logger.error(f"Unexpected Codex error: {type(e).__name__}: {e}")
+                raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("Codex call failed after all retries")
+
+    async def _parse_sse_stream(self, response: httpx.Response) -> str:
+        """
+        Parse Server-Sent Events (SSE) stream from Codex API.
+
+        Args:
+            response: HTTP response with SSE stream.
+
+        Returns:
+            Extracted text content from stream.
+        """
+        full_text = ""
+        event_type = None
+
+        async for line in response.aiter_lines():
+            if not line:
+                continue
+
+            # Track event type
+            if line.startswith("event: "):
+                event_type = line[7:]
+
+            # Parse data
+            elif line.startswith("data: "):
+                data_str = line[6:]
+                if data_str == "[DONE]":
+                    break
+
+                try:
+                    data = json.loads(data_str)
+
+                    # Extract content based on event type
+                    if event_type == "response.text.delta" and "delta" in data:
+                        full_text += data["delta"]
+                    elif event_type == "response.content_part.delta" and "delta" in data:
+                        full_text += data["delta"]
+                    # Check for item content
+                    elif "item" in data:
+                        item = data["item"]
+                        if "content" in item:
+                            content = item["content"]
+                            if isinstance(content, list):
+                                for part in content:
+                                    if isinstance(part, dict) and "text" in part:
+                                        full_text += part["text"]
+                            elif isinstance(content, str):
+                                full_text += content
+
+                except json.JSONDecodeError:
+                    # Skip malformed JSON events
+                    pass
+
+        return full_text
+
+    async def call_with_tools(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "tools",
+        max_retries: int = 5,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 30.0,
+        tool_choice: str | dict[str, Any] = "auto",
+    ) -> LLMToolCallResult:
+        """
+        Make API call with tool calling support.
+
+        Parses Codex SSE stream to extract tool calls from response.output_item.done events.
+        Tools are converted from OpenAI format to Codex format (flat structure at top level).
+
+        Args:
+            messages: List of message dicts. Can include tool results with role='tool'.
+            tools: List of tool definitions in OpenAI format.
+            max_completion_tokens: Maximum tokens in response.
+            temperature: Sampling temperature.
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            tool_choice: How to choose tools - "auto", "none", "required", or specific function.
+
+        Returns:
+            LLMToolCallResult with content and/or tool_calls.
+        """
+        start_time = time.time()
+
+        # Prepare system instructions
+        system_instruction = ""
+        user_messages = []
+
+        for msg in messages:
+            role = msg.get("role", "user")
+            content = msg.get("content", "")
+
+            if role == "system":
+                system_instruction += ("\n\n" + content) if system_instruction else content
+            elif role == "tool":
+                # Handle tool results
+                user_messages.append(
+                    {
+                        "type": "message",
+                        "role": "user",
+                        "content": f"Tool result: {content}",
+                    }
+                )
+            else:
+                user_messages.append(
+                    {
+                        "type": "message",
+                        "role": role,
+                        "content": content,
+                    }
+                )
+
+        # Convert tools to Codex format
+        # Codex expects tools with type and name/description/parameters at top level
+        codex_tools = []
+        for tool in tools:
+            func = tool.get("function", {})
+            codex_tools.append(
+                {
+                    "type": "function",
+                    "name": func.get("name", ""),
+                    "description": func.get("description", ""),
+                    "parameters": func.get("parameters", {}),
+                }
+            )
+
+        # gpt-5.2-codex only supports "detailed" reasoning summary
+        reasoning_summary = "detailed" if "5.2" in self.model else self.reasoning_summary
+
+        payload = {
+            "model": self.model,
+            "instructions": system_instruction,
+            "input": user_messages,
+            "tools": codex_tools,
+            "tool_choice": tool_choice,
+            "parallel_tool_calls": True,
+            "reasoning": {"summary": reasoning_summary},
+            "store": False,
+            "stream": True,
+            "include": ["reasoning.encrypted_content"],
+            "prompt_cache_key": str(uuid.uuid4()),
+        }
+
+        headers = {
+            "Authorization": f"Bearer {self.access_token}",
+            "Content-Type": "application/json",
+            "OpenAI-Account-ID": self.account_id,
+            "User-Agent": "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7)",
+            "Origin": "https://chatgpt.com",
+        }
+
+        url = f"{self.base_url}/codex/responses"
+
+        # Debug logging for troubleshooting
+        logger.debug(f"Codex tool call request: url={url}, model={payload['model']}, tools={len(codex_tools)}")
+
+        try:
+            response = await self._client.post(url, json=payload, headers=headers, timeout=120.0)
+
+            # Log response details on error
+            if response.status_code != 200:
+                logger.error(f"Codex API error {response.status_code}: {response.text[:500]}")
+
+            response.raise_for_status()
+
+            # Parse SSE for tool calls and content
+            content, tool_calls = await self._parse_sse_tool_stream(response)
+
+            duration = time.time() - start_time
+            metrics = get_metrics_collector()
+            metrics.record_llm_call(
+                provider=self.provider,
+                model=self.model,
+                scope=scope,
+                duration=duration,
+                input_tokens=0,
+                output_tokens=0,
+                success=True,
+            )
+
+            # Record OpenTelemetry span
+            from hindsight_api.tracing import get_span_recorder
+
+            span_recorder = get_span_recorder()
+            # Convert LLMToolCall objects to dicts for span recording
+            tool_calls_dict = (
+                [{"id": tc.id, "name": tc.name, "arguments": tc.arguments} for tc in tool_calls] if tool_calls else None
+            )
+            span_recorder.record_llm_call(
+                provider=self.provider,
+                model=self.model,
+                scope=scope,
+                messages=messages,
+                response_content=content,
+                input_tokens=0,  # Codex doesn't provide token counts
+                output_tokens=0,
+                duration=duration,
+                finish_reason="tool_calls" if tool_calls else "stop",
+                error=None,
+                tool_calls=tool_calls_dict,
+            )
+
+            return LLMToolCallResult(
+                content=content,
+                tool_calls=tool_calls,
+                finish_reason="tool_calls" if tool_calls else "stop",
+                input_tokens=0,
+                output_tokens=0,
+            )
+
+        except Exception as e:
+            logger.error(f"Codex tool call error: {e}")
+            raise
+
+    async def _parse_sse_tool_stream(self, response: httpx.Response) -> tuple[str | None, list[LLMToolCall]]:
+        """
+        Parse SSE stream for tool calls and content.
+
+        Returns:
+            Tuple of (content, tool_calls).
+        """
+        content = ""
+        tool_calls: list[LLMToolCall] = []
+        event_type = None
+
+        async for line in response.aiter_lines():
+            if not line:
+                continue
+
+            if line.startswith("event: "):
+                event_type = line[7:]
+
+            elif line.startswith("data: "):
+                data_str = line[6:]
+                if data_str == "[DONE]":
+                    break
+
+                try:
+                    data = json.loads(data_str)
+
+                    # Extract text content
+                    if event_type == "response.text.delta" and "delta" in data:
+                        content += data["delta"]
+
+                    # Extract completed tool calls from response.output_item.done
+                    elif event_type == "response.output_item.done":
+                        item = data.get("item", {})
+                        if item.get("type") == "function_call" and item.get("status") == "completed":
+                            tool_name = item.get("name", "")
+                            arguments_str = item.get("arguments", "{}")
+                            call_id = item.get("call_id", "")
+
+                            try:
+                                arguments = json.loads(arguments_str)
+                            except json.JSONDecodeError:
+                                logger.warning(f"Failed to parse tool arguments: {arguments_str}")
+                                arguments = {}
+
+                            tool_calls.append(
+                                LLMToolCall(
+                                    id=call_id,
+                                    name=tool_name,
+                                    arguments=arguments,
+                                )
+                            )
+
+                except json.JSONDecodeError as e:
+                    logger.warning(f"Failed to parse SSE data: {e}, data_str: {data_str[:200]}")
+
+        return content if content else None, tool_calls
+
+    async def cleanup(self) -> None:
+        """Clean up HTTP client."""
+        await self._client.aclose()

hindsight-api/hindsight_api/engine/providers/gemini_llm.py

@@ -0,0 +1,550 @@
+"""
+Google Gemini/VertexAI LLM provider.
+
+This provider supports both:
+1. Gemini API (api.generativeai.google.com) with API key authentication
+2. Vertex AI with service account or Application Default Credentials (ADC)
+"""
+
+import asyncio
+import json
+import logging
+import os
+import time
+from typing import Any
+
+from google import genai
+from google.genai import errors as genai_errors
+from google.genai import types as genai_types
+
+from hindsight_api.engine.llm_interface import LLMInterface, OutputTooLongError
+from hindsight_api.engine.response_models import LLMToolCall, LLMToolCallResult, TokenUsage
+from hindsight_api.metrics import get_metrics_collector
+
+logger = logging.getLogger(__name__)
+
+# Vertex AI imports (optional)
+try:
+    import google.auth
+    from google.oauth2 import service_account
+
+    VERTEXAI_AVAILABLE = True
+except ImportError:
+    VERTEXAI_AVAILABLE = False
+
+
+class GeminiLLM(LLMInterface):
+    """
+    LLM provider for Google Gemini and Vertex AI.
+
+    Supports:
+    - Gemini API: provider="gemini", requires api_key
+    - Vertex AI: provider="vertexai", requires project_id and region, uses ADC or service account
+    """
+
+    def __init__(
+        self,
+        provider: str,
+        api_key: str,
+        base_url: str,
+        model: str,
+        reasoning_effort: str = "low",
+        **kwargs: Any,
+    ):
+        """Initialize Gemini/VertexAI LLM provider."""
+        super().__init__(provider, api_key, base_url, model, reasoning_effort, **kwargs)
+
+        self._client = None
+        self._is_vertexai = self.provider == "vertexai"
+
+        if self._is_vertexai:
+            self._init_vertexai(**kwargs)
+        else:
+            self._init_gemini()
+
+    def _init_gemini(self) -> None:
+        """Initialize Gemini API client."""
+        if not self.api_key:
+            raise ValueError("Gemini provider requires api_key")
+
+        self._client = genai.Client(api_key=self.api_key)
+        logger.info(f"Gemini API: model={self.model}")
+
+    def _init_vertexai(self, **kwargs: Any) -> None:
+        """Initialize Vertex AI client with project, region, and credentials."""
+        # Extract Vertex AI config from kwargs
+        project_id = kwargs.get("vertexai_project_id")
+        region = kwargs.get("vertexai_region", "us-central1")
+        service_account_key = kwargs.get("vertexai_service_account_key")
+        credentials = kwargs.get("vertexai_credentials")  # Pre-loaded credentials object
+
+        if not project_id:
+            raise ValueError(
+                "HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID is required for Vertex AI provider. "
+                "Set it to your GCP project ID."
+            )
+
+        auth_method = "ADC"
+
+        # Use pre-loaded credentials if provided (passed from LLMProvider)
+        if credentials is not None:
+            auth_method = "service_account"
+        # Otherwise, load explicit service account credentials if path provided
+        elif service_account_key:
+            if not VERTEXAI_AVAILABLE:
+                raise ValueError(
+                    "Vertex AI service account auth requires 'google-auth' package. "
+                    "Install with: pip install google-auth"
+                )
+            credentials = service_account.Credentials.from_service_account_file(
+                service_account_key,
+                scopes=["https://www.googleapis.com/auth/cloud-platform"],
+            )
+            auth_method = "service_account"
+            logger.info(f"Vertex AI: Using service account key: {service_account_key}")
+
+        # Strip google/ prefix from model name — native SDK uses bare names
+        # e.g. "google/gemini-2.0-flash-lite-001" -> "gemini-2.0-flash-lite-001"
+        if self.model.startswith("google/"):
+            self.model = self.model[len("google/") :]
+
+        # Create Vertex AI client
+        client_kwargs: dict[str, Any] = {
+            "vertexai": True,
+            "project": project_id,
+            "location": region,
+        }
+        if credentials is not None:
+            client_kwargs["credentials"] = credentials
+
+        self._client = genai.Client(**client_kwargs)
+
+        logger.info(f"Vertex AI: project={project_id}, region={region}, model={self.model}, auth={auth_method}")
+
+    async def verify_connection(self) -> None:
+        """
+        Verify that the Gemini/VertexAI provider is configured correctly.
+
+        Raises:
+            RuntimeError: If the connection test fails.
+        """
+        try:
+            logger.info(f"Verifying {self.provider.upper()}: model={self.model}...")
+            await self.call(
+                messages=[{"role": "user", "content": "Say 'ok'"}],
+                max_completion_tokens=100,
+                max_retries=2,
+                initial_backoff=0.5,
+                max_backoff=2.0,
+                scope="verification",
+            )
+            logger.info(f"{self.provider.upper()} connection verified successfully")
+        except Exception as e:
+            raise RuntimeError(f"Failed to verify {self.provider.upper()} connection: {e}") from e
+
+    async def call(
+        self,
+        messages: list[dict[str, str]],
+        response_format: Any | None = None,
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "memory",
+        max_retries: int = 10,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 60.0,
+        skip_validation: bool = False,
+        strict_schema: bool = False,
+        return_usage: bool = False,
+    ) -> Any:
+        """
+        Make a Gemini/VertexAI API call with retry logic.
+
+        Args:
+            messages: List of message dicts with 'role' and 'content'.
+            response_format: Optional Pydantic model for structured output.
+            max_completion_tokens: Maximum tokens in response (not supported by Gemini).
+            temperature: Sampling temperature (0.0-2.0).
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            skip_validation: Return raw JSON without Pydantic validation.
+            strict_schema: Use strict JSON schema enforcement (not supported by Gemini).
+            return_usage: If True, return tuple (result, TokenUsage).
+
+        Returns:
+            If return_usage=False: Parsed response if response_format provided, else text.
+            If return_usage=True: Tuple of (result, TokenUsage).
+        """
+        start_time = time.time()
+
+        # Convert OpenAI-style messages to Gemini format
+        system_instruction = None
+        gemini_contents = []
+
+        for msg in messages:
+            role = msg.get("role", "user")
+            content = msg.get("content", "")
+
+            if role == "system":
+                if system_instruction:
+                    system_instruction += "\n\n" + content
+                else:
+                    system_instruction = content
+            elif role == "assistant":
+                gemini_contents.append(genai_types.Content(role="model", parts=[genai_types.Part(text=content)]))
+            else:
+                gemini_contents.append(genai_types.Content(role="user", parts=[genai_types.Part(text=content)]))
+
+        # Add JSON schema instruction if response_format is provided
+        if response_format is not None and hasattr(response_format, "model_json_schema"):
+            schema = response_format.model_json_schema()
+            schema_msg = f"\n\nYou must respond with valid JSON matching this schema:\n{json.dumps(schema, indent=2)}"
+            if system_instruction:
+                system_instruction += schema_msg
+            else:
+                system_instruction = schema_msg
+
+        # Build generation config
+        config_kwargs: dict[str, Any] = {}
+        if system_instruction:
+            config_kwargs["system_instruction"] = system_instruction
+        if response_format is not None:
+            config_kwargs["response_mime_type"] = "application/json"
+            config_kwargs["response_schema"] = response_format
+        if temperature is not None:
+            config_kwargs["temperature"] = temperature
+
+        generation_config = genai_types.GenerateContentConfig(**config_kwargs) if config_kwargs else None
+
+        last_exception = None
+
+        for attempt in range(max_retries + 1):
+            try:
+                response = await self._client.aio.models.generate_content(
+                    model=self.model,
+                    contents=gemini_contents,
+                    config=generation_config,
+                )
+
+                content = response.text
+
+                # Handle empty response
+                if content is None:
+                    block_reason = None
+                    if hasattr(response, "candidates") and response.candidates:
+                        candidate = response.candidates[0]
+                        if hasattr(candidate, "finish_reason"):
+                            block_reason = candidate.finish_reason
+
+                    if attempt < max_retries:
+                        logger.warning(f"Gemini returned empty response (reason: {block_reason}), retrying...")
+                        backoff = min(initial_backoff * (2**attempt), max_backoff)
+                        await asyncio.sleep(backoff)
+                        continue
+                    else:
+                        raise RuntimeError(f"Gemini returned empty response after {max_retries + 1} attempts")
+
+                # Parse structured output if requested
+                if response_format is not None:
+                    json_data = json.loads(content)
+                    if skip_validation:
+                        result = json_data
+                    else:
+                        result = response_format.model_validate(json_data)
+                else:
+                    result = content
+
+                # Extract token usage
+                input_tokens = 0
+                output_tokens = 0
+                if hasattr(response, "usage_metadata") and response.usage_metadata:
+                    usage = response.usage_metadata
+                    input_tokens = usage.prompt_token_count or 0
+                    output_tokens = usage.candidates_token_count or 0
+
+                # Record metrics
+                duration = time.time() - start_time
+                metrics = get_metrics_collector()
+                metrics.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    duration=duration,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    success=True,
+                )
+
+                # Record trace span
+                from hindsight_api.tracing import get_span_recorder
+
+                finish_reason = None
+                if hasattr(response, "candidates") and response.candidates:
+                    if hasattr(response.candidates[0], "finish_reason"):
+                        finish_reason = str(response.candidates[0].finish_reason)
+                span_recorder = get_span_recorder()
+                from hindsight_api.tracing import _serialize_for_span
+
+                span_recorder.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    messages=messages,
+                    response_content=_serialize_for_span(result),
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    duration=duration,
+                    finish_reason=finish_reason,
+                    error=None,
+                )
+
+                # Log slow calls
+                if duration > 10.0 and input_tokens > 0:
+                    logger.info(
+                        f"slow llm call: scope={scope}, model={self.provider}/{self.model}, "
+                        f"input_tokens={input_tokens}, output_tokens={output_tokens}, "
+                        f"time={duration:.3f}s"
+                    )
+
+                if return_usage:
+                    token_usage = TokenUsage(
+                        input_tokens=input_tokens,
+                        output_tokens=output_tokens,
+                        total_tokens=input_tokens + output_tokens,
+                    )
+                    return result, token_usage
+                return result
+
+            except json.JSONDecodeError as e:
+                last_exception = e
+                if attempt < max_retries:
+                    logger.warning("Gemini returned invalid JSON, retrying...")
+                    backoff = min(initial_backoff * (2**attempt), max_backoff)
+                    await asyncio.sleep(backoff)
+                    continue
+                else:
+                    logger.error(f"Gemini returned invalid JSON after {max_retries + 1} attempts")
+                    raise
+
+            except genai_errors.APIError as e:
+                # Fast fail on auth errors - these won't recover with retries
+                if e.code in (401, 403):
+                    logger.error(f"Gemini auth error (HTTP {e.code}), not retrying: {str(e)}")
+                    raise
+
+                # Retry on retryable errors (rate limits, server errors, client errors)
+                if e.code in (400, 429, 500, 502, 503, 504) or (e.code and e.code >= 500):
+                    last_exception = e
+                    if attempt < max_retries:
+                        backoff = min(initial_backoff * (2**attempt), max_backoff)
+                        jitter = backoff * 0.2 * (2 * (time.time() % 1) - 1)
+                        await asyncio.sleep(backoff + jitter)
+                    else:
+                        logger.error(f"Gemini API error after {max_retries + 1} attempts: {str(e)}")
+                        raise
+                else:
+                    logger.error(f"Gemini API error: {type(e).__name__}: {str(e)}")
+                    raise
+
+            except Exception as e:
+                logger.error(f"Unexpected error during Gemini call: {type(e).__name__}: {str(e)}")
+                raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("Gemini call failed after all retries")
+
+    async def call_with_tools(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "tools",
+        max_retries: int = 5,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 30.0,
+        tool_choice: str | dict[str, Any] = "auto",
+    ) -> LLMToolCallResult:
+        """
+        Make a Gemini/VertexAI API call with tool/function calling support.
+
+        Args:
+            messages: List of message dicts. Can include tool results with role='tool'.
+            tools: List of tool definitions in OpenAI format.
+            max_completion_tokens: Maximum tokens (not supported by Gemini).
+            temperature: Sampling temperature.
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            tool_choice: How to choose tools (Gemini uses "auto" only).
+
+        Returns:
+            LLMToolCallResult with content and/or tool_calls.
+        """
+        start_time = time.time()
+
+        # Convert tools to Gemini format
+        gemini_tools = []
+        for tool in tools:
+            func = tool.get("function", {})
+            gemini_tools.append(
+                genai_types.Tool(
+                    function_declarations=[
+                        genai_types.FunctionDeclaration(
+                            name=func.get("name", ""),
+                            description=func.get("description", ""),
+                            parameters=func.get("parameters"),
+                        )
+                    ]
+                )
+            )
+
+        # Convert messages
+        system_instruction = None
+        gemini_contents = []
+        for msg in messages:
+            role = msg.get("role", "user")
+            content = msg.get("content", "")
+
+            if role == "system":
+                system_instruction = (system_instruction + "\n\n" + content) if system_instruction else content
+            elif role == "tool":
+                # Gemini uses function_response
+                gemini_contents.append(
+                    genai_types.Content(
+                        role="user",
+                        parts=[
+                            genai_types.Part(
+                                function_response=genai_types.FunctionResponse(
+                                    name=msg.get("name", ""),
+                                    response={"result": content},
+                                )
+                            )
+                        ],
+                    )
+                )
+            elif role == "assistant":
+                gemini_contents.append(genai_types.Content(role="model", parts=[genai_types.Part(text=content)]))
+            else:
+                gemini_contents.append(genai_types.Content(role="user", parts=[genai_types.Part(text=content)]))
+
+        config_kwargs: dict[str, Any] = {"tools": gemini_tools}
+        if system_instruction:
+            config_kwargs["system_instruction"] = system_instruction
+        if temperature is not None:
+            config_kwargs["temperature"] = temperature
+
+        config = genai_types.GenerateContentConfig(**config_kwargs)
+
+        last_exception = None
+        for attempt in range(max_retries + 1):
+            try:
+                response = await self._client.aio.models.generate_content(
+                    model=self.model,
+                    contents=gemini_contents,
+                    config=config,
+                )
+
+                # Extract content and tool calls
+                content = None
+                tool_calls: list[LLMToolCall] = []
+
+                if response.candidates and response.candidates[0].content:
+                    parts = response.candidates[0].content.parts
+                    if parts:
+                        for part in parts:
+                            if hasattr(part, "text") and part.text:
+                                content = part.text
+                            if hasattr(part, "function_call") and part.function_call:
+                                fc = part.function_call
+                                tool_calls.append(
+                                    LLMToolCall(
+                                        id=f"gemini_{len(tool_calls)}",
+                                        name=fc.name,
+                                        arguments=dict(fc.args) if fc.args else {},
+                                    )
+                                )
+
+                finish_reason = "tool_calls" if tool_calls else "stop"
+
+                # Extract token usage
+                input_tokens = 0
+                output_tokens = 0
+                if response.usage_metadata:
+                    input_tokens = response.usage_metadata.prompt_token_count or 0
+                    output_tokens = response.usage_metadata.candidates_token_count or 0
+
+                # Record metrics
+                duration = time.time() - start_time
+                metrics = get_metrics_collector()
+                metrics.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    duration=duration,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    success=True,
+                )
+
+                # Record OpenTelemetry span
+                from hindsight_api.tracing import get_span_recorder
+
+                span_recorder = get_span_recorder()
+                # Convert LLMToolCall objects to dicts for span recording
+                tool_calls_dict = (
+                    [{"id": tc.id, "name": tc.name, "arguments": tc.arguments} for tc in tool_calls]
+                    if tool_calls
+                    else None
+                )
+                span_recorder.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    messages=messages,
+                    response_content=content,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    duration=duration,
+                    finish_reason=finish_reason,
+                    error=None,
+                    tool_calls=tool_calls_dict,
+                )
+
+                return LLMToolCallResult(
+                    content=content,
+                    tool_calls=tool_calls,
+                    finish_reason=finish_reason,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                )
+
+            except genai_errors.APIError as e:
+                # Fast fail on auth errors
+                if e.code in (401, 403):
+                    logger.error(f"Gemini auth error (HTTP {e.code}), not retrying: {str(e)}")
+                    raise
+
+                # Retry on retryable errors
+                last_exception = e
+                if attempt < max_retries:
+                    backoff = min(initial_backoff * (2**attempt), max_backoff)
+                    await asyncio.sleep(backoff)
+                    continue
+                raise
+
+            except Exception as e:
+                logger.error(f"Unexpected error during Gemini tool call: {type(e).__name__}: {str(e)}")
+                raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("Gemini tool call failed")
+
+    async def cleanup(self) -> None:
+        """Clean up resources (close connections, etc.)."""
+        # Gemini client doesn't require explicit cleanup
+        pass

hindsight-api/hindsight_api/engine/providers/mock_llm.py

@@ -0,0 +1,301 @@
+"""
+Mock LLM provider for testing.
+
+This provider allows tests to record LLM calls and return configurable mock responses
+without making actual API calls to external LLM services.
+"""
+
+import logging
+from typing import Any
+
+from ..llm_interface import LLMInterface
+from ..response_models import LLMToolCall, LLMToolCallResult, TokenUsage
+
+logger = logging.getLogger(__name__)
+
+
+class MockLLM(LLMInterface):
+    """
+    Mock LLM provider for testing.
+
+    This provider records all calls and returns configurable mock responses,
+    enabling tests to verify LLM interactions without making real API calls.
+
+    Example:
+        # Create mock provider
+        mock_llm = MockLLM(provider="mock", api_key="", base_url="", model="mock-model")
+
+        # Set mock response
+        mock_llm.set_mock_response({"answer": "test"})
+
+        # Make calls
+        result = await mock_llm.call(
+            messages=[{"role": "user", "content": "test"}],
+            response_format=MyResponseModel
+        )
+
+        # Verify calls
+        calls = mock_llm.get_mock_calls()
+        assert len(calls) == 1
+        assert calls[0]["scope"] == "memory"
+    """
+
+    def __init__(
+        self,
+        provider: str,
+        api_key: str,
+        base_url: str,
+        model: str,
+        reasoning_effort: str = "low",
+        **kwargs: Any,
+    ):
+        """
+        Initialize mock LLM provider.
+
+        Args:
+            provider: Provider name (should be "mock").
+            api_key: Not used for mock provider.
+            base_url: Not used for mock provider.
+            model: Model name for tracking.
+            reasoning_effort: Not used for mock provider.
+            **kwargs: Additional parameters (not used).
+        """
+        super().__init__(provider, api_key, base_url, model, reasoning_effort, **kwargs)
+
+        # Storage for test verification
+        self._mock_calls: list[dict] = []
+        self._mock_response: Any = None
+        self._mock_exception: Exception | None = None
+
+    async def verify_connection(self) -> None:
+        """
+        Verify mock provider (always succeeds).
+
+        Mock provider doesn't need connection verification since it doesn't
+        make real API calls.
+        """
+        logger.debug("Mock LLM: connection verification (always succeeds)")
+
+    async def call(
+        self,
+        messages: list[dict[str, str]],
+        response_format: Any | None = None,
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "memory",
+        max_retries: int = 10,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 60.0,
+        skip_validation: bool = False,
+        strict_schema: bool = False,
+        return_usage: bool = False,
+    ) -> Any:
+        """
+        Make a mock LLM API call.
+
+        Records the call for test verification and returns the configured mock response.
+
+        Args:
+            messages: List of message dicts with 'role' and 'content'.
+            response_format: Optional Pydantic model for structured output.
+            max_completion_tokens: Not used in mock.
+            temperature: Not used in mock.
+            scope: Scope identifier for tracking.
+            max_retries: Not used in mock.
+            initial_backoff: Not used in mock.
+            max_backoff: Not used in mock.
+            skip_validation: Return raw JSON without Pydantic validation.
+            strict_schema: Not used in mock.
+            return_usage: If True, return tuple (result, TokenUsage) instead of just result.
+
+        Returns:
+            If return_usage=False: Parsed response if response_format is provided, otherwise text content.
+            If return_usage=True: Tuple of (result, TokenUsage) with mock token counts.
+        """
+        # Record the call for test verification
+        call_record = {
+            "provider": self.provider,
+            "model": self.model,
+            "messages": messages,
+            "response_format": response_format.__name__
+            if response_format and hasattr(response_format, "__name__")
+            else str(response_format),
+            "scope": scope,
+        }
+        self._mock_calls.append(call_record)
+        logger.debug(f"Mock LLM call recorded: scope={scope}, model={self.model}")
+
+        # Raise mock exception if configured
+        if self._mock_exception is not None:
+            raise self._mock_exception
+
+        # Record trace span (minimal for mock provider)
+        from hindsight_api.tracing import get_span_recorder
+
+        span_recorder = get_span_recorder()
+        span_recorder.record_llm_call(
+            provider=self.provider,
+            model=self.model,
+            scope=scope,
+            messages=messages,
+            response_content="mock response",
+            input_tokens=10,
+            output_tokens=5,
+            duration=0.001,  # Mock calls are instant
+            finish_reason="stop",
+            error=None,
+        )
+
+        # Return mock response
+        if self._mock_response is not None:
+            result = self._mock_response
+        elif response_format is not None:
+            # Try to create a minimal valid instance of the response format
+            try:
+                # For Pydantic models, try to create with minimal valid data
+                result = {"mock": True}
+            except Exception:
+                result = {"mock": True}
+        else:
+            result = "mock response"
+
+        if return_usage:
+            token_usage = TokenUsage(input_tokens=10, output_tokens=5, total_tokens=15)
+            return result, token_usage
+        return result
+
+    async def call_with_tools(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "tools",
+        max_retries: int = 5,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 30.0,
+        tool_choice: str | dict[str, Any] = "auto",
+    ) -> LLMToolCallResult:
+        """
+        Make a mock LLM API call with tool/function calling support.
+
+        Records the call for test verification and returns the configured mock response.
+
+        Args:
+            messages: List of message dicts. Can include tool results with role='tool'.
+            tools: List of tool definitions in OpenAI format.
+            max_completion_tokens: Not used in mock.
+            temperature: Not used in mock.
+            scope: Scope identifier for tracking.
+            max_retries: Not used in mock.
+            initial_backoff: Not used in mock.
+            max_backoff: Not used in mock.
+            tool_choice: Not used in mock.
+
+        Returns:
+            LLMToolCallResult with content and/or tool_calls.
+        """
+        # Record the call for test verification
+        call_record = {
+            "provider": self.provider,
+            "model": self.model,
+            "messages": messages,
+            "tools": [t.get("function", {}).get("name") for t in tools],
+            "scope": scope,
+        }
+        self._mock_calls.append(call_record)
+
+        # Raise mock exception if configured
+        if self._mock_exception is not None:
+            raise self._mock_exception
+
+        # Record OpenTelemetry span
+        from hindsight_api.tracing import get_span_recorder
+
+        span_recorder = get_span_recorder()
+
+        if self._mock_response is not None:
+            if isinstance(self._mock_response, LLMToolCallResult):
+                result = self._mock_response
+            elif isinstance(self._mock_response, list):
+                # Allow setting just tool calls as a list
+                result = LLMToolCallResult(
+                    tool_calls=[
+                        LLMToolCall(id=f"mock_{i}", name=tc["name"], arguments=tc.get("arguments", {}))
+                        for i, tc in enumerate(self._mock_response)
+                    ],
+                    finish_reason="tool_calls",
+                )
+            else:
+                result = LLMToolCallResult(content="mock response", finish_reason="stop")
+        else:
+            result = LLMToolCallResult(content="mock response", finish_reason="stop")
+
+        # Record span with mock values
+        # Convert LLMToolCall objects to dicts for span recording
+        tool_calls_dict = (
+            [{"id": tc.id, "name": tc.name, "arguments": tc.arguments} for tc in result.tool_calls]
+            if result.tool_calls
+            else None
+        )
+        span_recorder.record_llm_call(
+            provider=self.provider,
+            model=self.model,
+            scope=scope,
+            messages=messages,
+            response_content=result.content,
+            input_tokens=10,  # Mock value
+            output_tokens=5,  # Mock value
+            duration=0.1,  # Mock value
+            finish_reason=result.finish_reason,
+            error=None,
+            tool_calls=tool_calls_dict,
+        )
+
+        return result
+
+    async def cleanup(self) -> None:
+        """Clean up resources (no-op for mock provider)."""
+        pass
+
+    def set_mock_response(self, response: Any) -> None:
+        """
+        Set the response to return from mock calls.
+
+        Args:
+            response: The response to return. Can be:
+                - A dict/Pydantic model for regular calls
+                - An LLMToolCallResult for tool calls
+                - A list of tool call dicts for tool calls
+                - Any other value to return as-is
+        """
+        self._mock_response = response
+
+    def set_mock_exception(self, exception: Exception) -> None:
+        """
+        Set an exception to raise from mock calls.
+
+        Args:
+            exception: The exception to raise on the next call.
+                      After raising, the exception is cleared.
+        """
+        self._mock_exception = exception
+
+    def get_mock_calls(self) -> list[dict]:
+        """
+        Get the list of recorded mock calls.
+
+        Returns:
+            List of call records, each containing:
+                - provider: Provider name
+                - model: Model name
+                - messages: Messages sent
+                - response_format/tools: Format or tools used
+                - scope: Call scope
+        """
+        return self._mock_calls
+
+    def clear_mock_calls(self) -> None:
+        """Clear the recorded mock calls and any set exception."""
+        self._mock_calls = []
+        self._mock_exception = None

hindsight-api/hindsight_api/engine/providers/openai_compatible_llm.py

@@ -0,0 +1,924 @@
+"""
+OpenAI-compatible LLM provider supporting OpenAI, Groq, Ollama, and LMStudio.
+
+This provider handles all OpenAI API-compatible models including:
+- OpenAI: GPT-4, GPT-4o, GPT-5, o1, o3 (reasoning models)
+- Groq: Fast inference with seed control and service tiers
+- Ollama: Local models with native streaming API support
+- LMStudio: Local models with OpenAI-compatible API
+
+Features:
+- Reasoning models with extended thinking (o1, o3, GPT-5 families)
+- Strict JSON schema enforcement (OpenAI)
+- Provider-specific parameters (Groq seed, service tier)
+- Native Ollama streaming for better structured output
+- Automatic token limit handling per model family
+"""
+
+import asyncio
+import io
+import json
+import logging
+import os
+import re
+import time
+from typing import Any
+
+import httpx
+from openai import APIConnectionError, APIStatusError, AsyncOpenAI, LengthFinishReasonError
+
+from hindsight_api.config import DEFAULT_LLM_TIMEOUT, ENV_LLM_TIMEOUT
+from hindsight_api.engine.llm_interface import LLMInterface, OutputTooLongError
+from hindsight_api.engine.response_models import LLMToolCall, LLMToolCallResult, TokenUsage
+from hindsight_api.metrics import get_metrics_collector
+
+logger = logging.getLogger(__name__)
+
+# Seed applied to every Groq request for deterministic behavior
+DEFAULT_LLM_SEED = 4242
+
+
+class OpenAICompatibleLLM(LLMInterface):
+    """
+    LLM provider for OpenAI-compatible APIs.
+
+    Supports:
+    - OpenAI: Standard models (GPT-4, GPT-4o) and reasoning models (o1, o3, GPT-5)
+    - Groq: Fast inference with seed control and service tiers
+    - Ollama: Local models with native streaming API for better structured output
+    - LMStudio: Local models with OpenAI-compatible API
+    """
+
+    def __init__(
+        self,
+        provider: str,
+        api_key: str,
+        base_url: str,
+        model: str,
+        reasoning_effort: str = "low",
+        timeout: float | None = None,
+        groq_service_tier: str | None = None,
+        **kwargs: Any,
+    ):
+        """
+        Initialize OpenAI-compatible LLM provider.
+
+        Args:
+            provider: Provider name ("openai", "groq", "ollama", "lmstudio").
+            api_key: API key (optional for ollama/lmstudio).
+            base_url: Base URL for the API (uses defaults for groq/ollama/lmstudio if empty).
+            model: Model name.
+            reasoning_effort: Reasoning effort level for supported models ("low", "medium", "high").
+            timeout: Request timeout in seconds (uses env var or 300s default).
+            groq_service_tier: Groq service tier ("on_demand", "flex", "auto").
+            **kwargs: Additional provider-specific parameters.
+        """
+        super().__init__(provider, api_key, base_url, model, reasoning_effort, **kwargs)
+
+        # Validate provider
+        valid_providers = ["openai", "groq", "ollama", "lmstudio"]
+        if self.provider not in valid_providers:
+            raise ValueError(f"OpenAICompatibleLLM only supports: {', '.join(valid_providers)}. Got: {self.provider}")
+
+        # Set default base URLs
+        if not self.base_url:
+            if self.provider == "groq":
+                self.base_url = "https://api.groq.com/openai/v1"
+            elif self.provider == "ollama":
+                self.base_url = "http://localhost:11434/v1"
+            elif self.provider == "lmstudio":
+                self.base_url = "http://localhost:1234/v1"
+
+        # For ollama/lmstudio, use dummy key if not provided
+        if self.provider in ("ollama", "lmstudio") and not self.api_key:
+            self.api_key = "local"
+
+        # Validate API key for cloud providers
+        if self.provider in ("openai", "groq") and not self.api_key:
+            raise ValueError(f"API key is required for {self.provider}")
+
+        # Service tier configuration (from config, not env vars)
+        self.groq_service_tier = groq_service_tier
+        self.openai_service_tier = kwargs.get("openai_service_tier")
+
+        # Get timeout config
+        self.timeout = timeout or float(os.getenv(ENV_LLM_TIMEOUT, str(DEFAULT_LLM_TIMEOUT)))
+
+        # Create OpenAI client
+        client_kwargs: dict[str, Any] = {"api_key": self.api_key, "max_retries": 0}
+        if self.base_url:
+            client_kwargs["base_url"] = self.base_url
+        if self.timeout:
+            client_kwargs["timeout"] = self.timeout
+
+        self._client = AsyncOpenAI(**client_kwargs)
+        logger.info(
+            f"OpenAI-compatible client initialized: provider={self.provider}, model={self.model}, "
+            f"base_url={self.base_url or 'default'}"
+        )
+
+    async def verify_connection(self) -> None:
+        """
+        Verify that the provider is configured correctly by making a simple test call.
+
+        Raises:
+            RuntimeError: If the connection test fails.
+        """
+        try:
+            logger.info(f"Verifying connection: {self.provider}/{self.model}")
+            await self.call(
+                messages=[{"role": "user", "content": "Say 'ok'"}],
+                max_completion_tokens=100,
+                max_retries=2,
+                initial_backoff=0.5,
+                max_backoff=2.0,
+                scope="verification",
+            )
+            logger.info(f"Connection verified: {self.provider}/{self.model}")
+        except Exception as e:
+            raise RuntimeError(f"Connection verification failed for {self.provider}/{self.model}: {e}") from e
+
+    def _supports_reasoning_model(self) -> bool:
+        """Check if the current model is a reasoning model (o1, o3, GPT-5, DeepSeek)."""
+        model_lower = self.model.lower()
+        return any(x in model_lower for x in ["gpt-5", "o1", "o3", "deepseek"])
+
+    def _get_max_reasoning_tokens(self) -> int | None:
+        """Get max reasoning tokens for reasoning models."""
+        model_lower = self.model.lower()
+
+        # GPT-4 and GPT-4.1 models have different caps
+        if any(x in model_lower for x in ["gpt-4.1", "gpt-4-"]):
+            return 32000
+        elif "gpt-4o" in model_lower:
+            return 16384
+
+        return None
+
+    async def call(
+        self,
+        messages: list[dict[str, str]],
+        response_format: Any | None = None,
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "memory",
+        max_retries: int = 10,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 60.0,
+        skip_validation: bool = False,
+        strict_schema: bool = False,
+        return_usage: bool = False,
+    ) -> Any:
+        """
+        Make an LLM API call with retry logic.
+
+        Args:
+            messages: List of message dicts with 'role' and 'content'.
+            response_format: Optional Pydantic model for structured output.
+            max_completion_tokens: Maximum tokens in response.
+            temperature: Sampling temperature (0.0-2.0).
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            skip_validation: Return raw JSON without Pydantic validation.
+            strict_schema: Use strict JSON schema enforcement (OpenAI only).
+            return_usage: If True, return tuple (result, TokenUsage) instead of just result.
+
+        Returns:
+            If return_usage=False: Parsed response if response_format is provided, otherwise text content.
+            If return_usage=True: Tuple of (result, TokenUsage) with token counts.
+
+        Raises:
+            OutputTooLongError: If output exceeds token limits.
+            Exception: Re-raises API errors after retries exhausted.
+        """
+        # Handle Ollama with native API for structured output (better schema enforcement)
+        if self.provider == "ollama" and response_format is not None:
+            return await self._call_ollama_native(
+                messages=messages,
+                response_format=response_format,
+                max_completion_tokens=max_completion_tokens,
+                temperature=temperature,
+                max_retries=max_retries,
+                initial_backoff=initial_backoff,
+                max_backoff=max_backoff,
+                skip_validation=skip_validation,
+                scope=scope,
+                return_usage=return_usage,
+            )
+
+        start_time = time.time()
+
+        # Build call parameters
+        call_params: dict[str, Any] = {
+            "model": self.model,
+            "messages": messages,
+        }
+
+        # Check if model supports reasoning parameter
+        is_reasoning_model = self._supports_reasoning_model()
+
+        # Apply model-specific token limits
+        if max_completion_tokens is not None:
+            max_tokens_cap = self._get_max_reasoning_tokens()
+            if max_tokens_cap and max_completion_tokens > max_tokens_cap:
+                max_completion_tokens = max_tokens_cap
+            # For reasoning models, enforce minimum to ensure space for reasoning + output
+            if is_reasoning_model and max_completion_tokens < 16000:
+                max_completion_tokens = 16000
+            call_params["max_completion_tokens"] = max_completion_tokens
+
+        # Temperature - reasoning models don't support custom temperature
+        if temperature is not None and not is_reasoning_model:
+            call_params["temperature"] = temperature
+
+        # Set reasoning_effort for reasoning models
+        if is_reasoning_model:
+            call_params["reasoning_effort"] = self.reasoning_effort
+
+        # Provider-specific parameters
+        if self.provider == "groq":
+            call_params["seed"] = DEFAULT_LLM_SEED
+            extra_body: dict[str, Any] = {}
+            # Add service_tier if configured
+            if self.groq_service_tier:
+                extra_body["service_tier"] = self.groq_service_tier
+            # Add reasoning parameters for reasoning models
+            if is_reasoning_model:
+                extra_body["include_reasoning"] = False
+            if extra_body:
+                call_params["extra_body"] = extra_body
+
+        # Prepare response format ONCE before retry loop
+        if response_format is not None:
+            schema = None
+            if hasattr(response_format, "model_json_schema"):
+                schema = response_format.model_json_schema()
+
+            if strict_schema and schema is not None:
+                # Use OpenAI's strict JSON schema enforcement
+                call_params["response_format"] = {
+                    "type": "json_schema",
+                    "json_schema": {
+                        "name": "response",
+                        "strict": True,
+                        "schema": schema,
+                    },
+                }
+            else:
+                # Soft enforcement: add schema to prompt and use json_object mode
+                if schema is not None:
+                    schema_msg = (
+                        f"\n\nYou must respond with valid JSON matching this schema:\n{json.dumps(schema, indent=2)}"
+                    )
+
+                    if call_params["messages"] and call_params["messages"][0].get("role") == "system":
+                        first_msg = call_params["messages"][0]
+                        if isinstance(first_msg, dict) and isinstance(first_msg.get("content"), str):
+                            first_msg["content"] += schema_msg
+                    elif call_params["messages"]:
+                        first_msg = call_params["messages"][0]
+                        if isinstance(first_msg, dict) and isinstance(first_msg.get("content"), str):
+                            first_msg["content"] = schema_msg + "\n\n" + first_msg["content"]
+                if self.provider not in ("lmstudio", "ollama"):
+                    # LM Studio and Ollama don't support json_object response format reliably
+                    call_params["response_format"] = {"type": "json_object"}
+
+        last_exception = None
+
+        for attempt in range(max_retries + 1):
+            try:
+                if response_format is not None:
+                    response = await self._client.chat.completions.create(**call_params)
+
+                    content = response.choices[0].message.content
+
+                    # Strip reasoning model thinking tags
+                    # Supports: <think>, <thinking>, <reasoning>, |startthink|/|endthink|
+                    if content:
+                        original_len = len(content)
+                        content = re.sub(r"<think>.*?</think>", "", content, flags=re.DOTALL)
+                        content = re.sub(r"<thinking>.*?</thinking>", "", content, flags=re.DOTALL)
+                        content = re.sub(r"<reasoning>.*?</reasoning>", "", content, flags=re.DOTALL)
+                        content = re.sub(r"\|startthink\|.*?\|endthink\|", "", content, flags=re.DOTALL)
+                        content = content.strip()
+                        if len(content) < original_len:
+                            logger.debug(f"Stripped {original_len - len(content)} chars of reasoning tokens")
+
+                    # For local models, they may wrap JSON in markdown code blocks
+                    if self.provider in ("lmstudio", "ollama"):
+                        clean_content = content
+                        if "```json" in content:
+                            clean_content = content.split("```json")[1].split("```")[0].strip()
+                        elif "```" in content:
+                            clean_content = content.split("```")[1].split("```")[0].strip()
+                        try:
+                            json_data = json.loads(clean_content)
+                        except json.JSONDecodeError:
+                            # Fallback to parsing raw content
+                            json_data = json.loads(content)
+                    else:
+                        # Log raw LLM response for debugging JSON parse issues
+                        try:
+                            json_data = json.loads(content)
+                        except json.JSONDecodeError as json_err:
+                            # Truncate content for logging
+                            content_preview = content[:500] if content else "<empty>"
+                            if content and len(content) > 700:
+                                content_preview = f"{content[:500]}...TRUNCATED...{content[-200:]}"
+                            logger.warning(
+                                f"JSON parse error from LLM response (attempt {attempt + 1}/{max_retries + 1}): {json_err}\n"
+                                f"  Model: {self.provider}/{self.model}\n"
+                                f"  Content length: {len(content) if content else 0} chars\n"
+                                f"  Content preview: {content_preview!r}\n"
+                                f"  Finish reason: {response.choices[0].finish_reason if response.choices else 'unknown'}"
+                            )
+                            # Retry on JSON parse errors
+                            if attempt < max_retries:
+                                backoff = min(initial_backoff * (2**attempt), max_backoff)
+                                await asyncio.sleep(backoff)
+                                last_exception = json_err
+                                continue
+                            else:
+                                logger.error(f"JSON parse error after {max_retries + 1} attempts, giving up")
+                                raise
+
+                    if skip_validation:
+                        result = json_data
+                    else:
+                        result = response_format.model_validate(json_data)
+                else:
+                    response = await self._client.chat.completions.create(**call_params)
+                    result = response.choices[0].message.content
+
+                # Record token usage metrics
+                duration = time.time() - start_time
+                usage = response.usage
+                input_tokens = usage.prompt_tokens or 0 if usage else 0
+                output_tokens = usage.completion_tokens or 0 if usage else 0
+                total_tokens = usage.total_tokens or 0 if usage else 0
+
+                # Record LLM metrics
+                metrics = get_metrics_collector()
+                metrics.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    duration=duration,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    success=True,
+                )
+
+                # Record trace span
+                from hindsight_api.tracing import _serialize_for_span, get_span_recorder
+
+                finish_reason = response.choices[0].finish_reason if response.choices else None
+                span_recorder = get_span_recorder()
+                span_recorder.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    messages=messages,
+                    response_content=_serialize_for_span(result),
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    duration=duration,
+                    finish_reason=finish_reason,
+                    error=None,
+                )
+
+                # Log slow calls
+                if duration > 10.0 and usage:
+                    ratio = max(1, output_tokens) / max(1, input_tokens)
+                    cached_tokens = 0
+                    if hasattr(usage, "prompt_tokens_details") and usage.prompt_tokens_details:
+                        cached_tokens = getattr(usage.prompt_tokens_details, "cached_tokens", 0) or 0
+                    cache_info = f", cached_tokens={cached_tokens}" if cached_tokens > 0 else ""
+                    logger.info(
+                        f"slow llm call: scope={scope}, model={self.provider}/{self.model}, "
+                        f"input_tokens={input_tokens}, output_tokens={output_tokens}, "
+                        f"total_tokens={total_tokens}{cache_info}, time={duration:.3f}s, ratio out/in={ratio:.2f}"
+                    )
+
+                if return_usage:
+                    token_usage = TokenUsage(
+                        input_tokens=input_tokens,
+                        output_tokens=output_tokens,
+                        total_tokens=total_tokens,
+                    )
+                    return result, token_usage
+                return result
+
+            except LengthFinishReasonError as e:
+                logger.warning(f"LLM output exceeded token limits: {str(e)}")
+                raise OutputTooLongError(
+                    "LLM output exceeded token limits. Input may need to be split into smaller chunks."
+                ) from e
+
+            except APIConnectionError as e:
+                last_exception = e
+                status_code = getattr(e, "status_code", None) or getattr(
+                    getattr(e, "response", None), "status_code", None
+                )
+                logger.warning(f"APIConnectionError (HTTP {status_code}), attempt {attempt + 1}: {str(e)[:200]}")
+                if attempt < max_retries:
+                    backoff = min(initial_backoff * (2**attempt), max_backoff)
+                    await asyncio.sleep(backoff)
+                    continue
+                else:
+                    logger.error(f"Connection error after {max_retries + 1} attempts: {str(e)}")
+                    raise
+
+            except APIStatusError as e:
+                # Fast fail only on 401 (unauthorized) and 403 (forbidden)
+                if e.status_code in (401, 403):
+                    logger.error(f"Auth error (HTTP {e.status_code}), not retrying: {str(e)}")
+                    raise
+
+                # Handle tool_use_failed error - model outputted in tool call format
+                if e.status_code == 400 and response_format is not None:
+                    try:
+                        error_body = e.body if hasattr(e, "body") else {}
+                        if isinstance(error_body, dict):
+                            error_info: dict[str, Any] = error_body.get("error") or {}
+                            if error_info.get("code") == "tool_use_failed":
+                                failed_gen = error_info.get("failed_generation", "")
+                                if failed_gen:
+                                    # Parse tool call format and convert to expected format
+                                    tool_call = json.loads(failed_gen)
+                                    tool_name = tool_call.get("name", "")
+                                    tool_args = tool_call.get("arguments", {})
+                                    converted = {"actions": [{"tool": tool_name, **tool_args}]}
+                                    if skip_validation:
+                                        result = converted
+                                    else:
+                                        result = response_format.model_validate(converted)
+
+                                    # Record metrics
+                                    duration = time.time() - start_time
+                                    metrics = get_metrics_collector()
+                                    metrics.record_llm_call(
+                                        provider=self.provider,
+                                        model=self.model,
+                                        scope=scope,
+                                        duration=duration,
+                                        input_tokens=0,
+                                        output_tokens=0,
+                                        success=True,
+                                    )
+                                    if return_usage:
+                                        return result, TokenUsage(input_tokens=0, output_tokens=0, total_tokens=0)
+                                    return result
+                    except (json.JSONDecodeError, KeyError, TypeError):
+                        pass  # Failed to parse tool_use_failed, continue with normal retry
+
+                last_exception = e
+                if attempt < max_retries:
+                    backoff = min(initial_backoff * (2**attempt), max_backoff)
+                    jitter = backoff * 0.2 * (2 * (time.time() % 1) - 1)
+                    sleep_time = backoff + jitter
+                    await asyncio.sleep(sleep_time)
+                else:
+                    logger.error(f"API error after {max_retries + 1} attempts: {str(e)}")
+                    raise
+
+            except Exception:
+                raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("LLM call failed after all retries with no exception captured")
+
+    async def call_with_tools(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "tools",
+        max_retries: int = 5,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 30.0,
+        tool_choice: str | dict[str, Any] = "auto",
+    ) -> LLMToolCallResult:
+        """
+        Make an LLM API call with tool/function calling support.
+
+        Args:
+            messages: List of message dicts. Can include tool results with role='tool'.
+            tools: List of tool definitions in OpenAI format.
+            max_completion_tokens: Maximum tokens in response.
+            temperature: Sampling temperature (0.0-2.0).
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            tool_choice: How to choose tools - "auto", "none", "required", or specific function.
+
+        Returns:
+            LLMToolCallResult with content and/or tool_calls.
+        """
+        start_time = time.time()
+
+        # Build call parameters
+        call_params: dict[str, Any] = {
+            "model": self.model,
+            "messages": messages,
+            "tools": tools,
+            "tool_choice": tool_choice,
+        }
+
+        if max_completion_tokens is not None:
+            call_params["max_completion_tokens"] = max_completion_tokens
+        if temperature is not None:
+            call_params["temperature"] = temperature
+
+        # Provider-specific parameters
+        if self.provider == "groq":
+            call_params["seed"] = DEFAULT_LLM_SEED
+
+        last_exception = None
+
+        for attempt in range(max_retries + 1):
+            try:
+                response = await self._client.chat.completions.create(**call_params)
+
+                message = response.choices[0].message
+                finish_reason = response.choices[0].finish_reason
+
+                # Extract tool calls if present
+                tool_calls: list[LLMToolCall] = []
+                if message.tool_calls:
+                    for tc in message.tool_calls:
+                        try:
+                            args = json.loads(tc.function.arguments) if tc.function.arguments else {}
+                        except json.JSONDecodeError:
+                            args = {"_raw": tc.function.arguments}
+                        tool_calls.append(LLMToolCall(id=tc.id, name=tc.function.name, arguments=args))
+
+                content = message.content
+
+                # Record metrics
+                duration = time.time() - start_time
+                usage = response.usage
+                input_tokens = usage.prompt_tokens or 0 if usage else 0
+                output_tokens = usage.completion_tokens or 0 if usage else 0
+
+                metrics = get_metrics_collector()
+                metrics.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    duration=duration,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    success=True,
+                )
+
+                # Record OpenTelemetry span
+                from hindsight_api.tracing import get_span_recorder
+
+                span_recorder = get_span_recorder()
+                # Convert LLMToolCall objects to dicts for span recording
+                tool_calls_dict = (
+                    [{"id": tc.id, "name": tc.name, "arguments": tc.arguments} for tc in tool_calls]
+                    if tool_calls
+                    else None
+                )
+                span_recorder.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    messages=messages,
+                    response_content=content,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    duration=duration,
+                    finish_reason=finish_reason,
+                    error=None,
+                    tool_calls=tool_calls_dict,
+                )
+
+                return LLMToolCallResult(
+                    content=content,
+                    tool_calls=tool_calls,
+                    finish_reason=finish_reason,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                )
+
+            except APIConnectionError as e:
+                last_exception = e
+                if attempt < max_retries:
+                    await asyncio.sleep(min(initial_backoff * (2**attempt), max_backoff))
+                    continue
+                raise
+
+            except APIStatusError as e:
+                if e.status_code in (401, 403):
+                    raise
+                last_exception = e
+                if attempt < max_retries:
+                    await asyncio.sleep(min(initial_backoff * (2**attempt), max_backoff))
+                    continue
+                raise
+
+            except Exception:
+                raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("Tool call failed after all retries")
+
+    async def _call_ollama_native(
+        self,
+        messages: list[dict[str, str]],
+        response_format: Any,
+        max_completion_tokens: int | None,
+        temperature: float | None,
+        max_retries: int,
+        initial_backoff: float,
+        max_backoff: float,
+        skip_validation: bool,
+        scope: str = "memory",
+        return_usage: bool = False,
+    ) -> Any:
+        """
+        Call Ollama using native API with JSON schema enforcement.
+
+        Ollama's native API supports passing a full JSON schema in the 'format' parameter,
+        which provides better structured output control than the OpenAI-compatible API.
+        """
+        start_time = time.time()
+
+        # Get the JSON schema from the Pydantic model
+        schema = response_format.model_json_schema() if hasattr(response_format, "model_json_schema") else None
+
+        # Build the base URL for Ollama's native API
+        # Default OpenAI-compatible URL is http://localhost:11434/v1
+        # Native API is at http://localhost:11434/api/chat
+        base_url = self.base_url or "http://localhost:11434/v1"
+        if base_url.endswith("/v1"):
+            native_url = base_url[:-3] + "/api/chat"
+        else:
+            native_url = base_url.rstrip("/") + "/api/chat"
+
+        # Build request payload
+        payload: dict[str, Any] = {
+            "model": self.model,
+            "messages": messages,
+            "stream": False,
+        }
+
+        # Add schema as format parameter for structured output
+        if schema:
+            payload["format"] = schema
+
+        # Add optional parameters with optimized defaults for Ollama
+        options: dict[str, Any] = {
+            "num_ctx": 16384,  # 16k context window for larger prompts
+            "num_batch": 512,  # Optimal batch size for prompt processing
+        }
+        if max_completion_tokens:
+            options["num_predict"] = max_completion_tokens
+        if temperature is not None:
+            options["temperature"] = temperature
+        payload["options"] = options
+
+        last_exception = None
+
+        async with httpx.AsyncClient(timeout=300.0) as client:
+            for attempt in range(max_retries + 1):
+                try:
+                    response = await client.post(native_url, json=payload)
+                    response.raise_for_status()
+
+                    result = response.json()
+                    content = result.get("message", {}).get("content", "")
+
+                    # Parse JSON response
+                    try:
+                        json_data = json.loads(content)
+                    except json.JSONDecodeError as json_err:
+                        content_preview = content[:500] if content else "<empty>"
+                        if content and len(content) > 700:
+                            content_preview = f"{content[:500]}...TRUNCATED...{content[-200:]}"
+                        logger.warning(
+                            f"Ollama JSON parse error (attempt {attempt + 1}/{max_retries + 1}): {json_err}\n"
+                            f"  Model: ollama/{self.model}\n"
+                            f"  Content length: {len(content) if content else 0} chars\n"
+                            f"  Content preview: {content_preview!r}"
+                        )
+                        if attempt < max_retries:
+                            backoff = min(initial_backoff * (2**attempt), max_backoff)
+                            await asyncio.sleep(backoff)
+                            last_exception = json_err
+                            continue
+                        else:
+                            raise
+
+                    # Extract token usage from Ollama response
+                    duration = time.time() - start_time
+                    input_tokens = result.get("prompt_eval_count", 0) or 0
+                    output_tokens = result.get("eval_count", 0) or 0
+                    total_tokens = input_tokens + output_tokens
+
+                    # Record LLM metrics
+                    metrics = get_metrics_collector()
+                    metrics.record_llm_call(
+                        provider=self.provider,
+                        model=self.model,
+                        scope=scope,
+                        duration=duration,
+                        input_tokens=input_tokens,
+                        output_tokens=output_tokens,
+                        success=True,
+                    )
+
+                    # Validate against Pydantic model or return raw JSON
+                    if skip_validation:
+                        validated_result = json_data
+                    else:
+                        validated_result = response_format.model_validate(json_data)
+
+                    if return_usage:
+                        token_usage = TokenUsage(
+                            input_tokens=input_tokens,
+                            output_tokens=output_tokens,
+                            total_tokens=total_tokens,
+                        )
+                        return validated_result, token_usage
+                    return validated_result
+
+                except httpx.HTTPStatusError as e:
+                    last_exception = e
+                    if attempt < max_retries:
+                        logger.warning(
+                            f"Ollama HTTP error (attempt {attempt + 1}/{max_retries + 1}): {e.response.status_code}"
+                        )
+                        backoff = min(initial_backoff * (2**attempt), max_backoff)
+                        await asyncio.sleep(backoff)
+                        continue
+                    else:
+                        logger.error(f"Ollama HTTP error after {max_retries + 1} attempts: {e}")
+                        raise
+
+                except httpx.RequestError as e:
+                    last_exception = e
+                    if attempt < max_retries:
+                        logger.warning(f"Ollama connection error (attempt {attempt + 1}/{max_retries + 1}): {e}")
+                        backoff = min(initial_backoff * (2**attempt), max_backoff)
+                        await asyncio.sleep(backoff)
+                        continue
+                    else:
+                        logger.error(f"Ollama connection error after {max_retries + 1} attempts: {e}")
+                        raise
+
+                except Exception as e:
+                    logger.error(f"Unexpected error during Ollama call: {type(e).__name__}: {e}")
+                    raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("Ollama call failed after all retries")
+
+    async def supports_batch_api(self) -> bool:
+        """Check if this provider supports batch API operations."""
+        # Only OpenAI and Groq support batch API
+        return self.provider in ("openai", "groq")
+
+    async def submit_batch(
+        self,
+        requests: list[dict[str, Any]],
+        endpoint: str = "/v1/chat/completions",
+        completion_window: str = "24h",
+    ) -> dict[str, Any]:
+        """
+        Submit a batch of requests to OpenAI/Groq Batch API.
+
+        Args:
+            requests: List of request dicts with custom_id, method, url, body
+            endpoint: API endpoint (e.g., "/v1/chat/completions")
+            completion_window: Completion window (e.g., "24h")
+
+        Returns:
+            Dict with batch metadata including batch_id
+
+        Raises:
+            NotImplementedError: If provider doesn't support batch API
+        """
+        if not await self.supports_batch_api():
+            raise NotImplementedError(f"Batch API not supported for provider: {self.provider}")
+
+        logger.info(f"Submitting batch with {len(requests)} requests to {self.provider}")
+
+        # Format requests as JSONL
+        jsonl_content = "\n".join(json.dumps(req) for req in requests)
+
+        # Upload file to provider (wrap in BytesIO with filename)
+        file_bytes = io.BytesIO(jsonl_content.encode("utf-8"))
+        file_bytes.name = "batch_input.jsonl"  # OpenAI SDK needs a filename
+
+        file_response = await self._client.files.create(
+            file=file_bytes,
+            purpose="batch",
+        )
+
+        logger.debug(f"Uploaded batch file: {file_response.id}")
+
+        # Create batch
+        batch_response = await self._client.batches.create(
+            input_file_id=file_response.id,
+            endpoint=endpoint,
+            completion_window=completion_window,
+        )
+
+        logger.info(f"Batch submitted: {batch_response.id}, status={batch_response.status}")
+
+        return {
+            "batch_id": batch_response.id,
+            "status": batch_response.status,
+            "input_file_id": file_response.id,
+            "created_at": batch_response.created_at,
+            "request_count": len(requests),
+        }
+
+    async def get_batch_status(self, batch_id: str) -> dict[str, Any]:
+        """
+        Get the status of a batch job.
+
+        Args:
+            batch_id: Batch identifier
+
+        Returns:
+            Dict with status info (batch_id, status, completed_at, etc.)
+        """
+        if not await self.supports_batch_api():
+            raise NotImplementedError(f"Batch API not supported for provider: {self.provider}")
+
+        batch = await self._client.batches.retrieve(batch_id)
+
+        result = {
+            "batch_id": batch.id,
+            "status": batch.status,
+            "created_at": batch.created_at,
+            "request_counts": {
+                "total": batch.request_counts.total if batch.request_counts else 0,
+                "completed": batch.request_counts.completed if batch.request_counts else 0,
+                "failed": batch.request_counts.failed if batch.request_counts else 0,
+            },
+        }
+
+        if batch.completed_at:
+            result["completed_at"] = batch.completed_at
+        if batch.output_file_id:
+            result["output_file_id"] = batch.output_file_id
+        if batch.error_file_id:
+            result["error_file_id"] = batch.error_file_id
+        if batch.errors:
+            result["errors"] = batch.errors
+
+        return result
+
+    async def retrieve_batch_results(self, batch_id: str) -> list[dict[str, Any]]:
+        """
+        Retrieve completed batch results.
+
+        Args:
+            batch_id: Batch identifier
+
+        Returns:
+            List of result dicts (one per request, matched by custom_id)
+        """
+        if not await self.supports_batch_api():
+            raise NotImplementedError(f"Batch API not supported for provider: {self.provider}")
+
+        # Get batch status
+        batch = await self._client.batches.retrieve(batch_id)
+
+        if batch.status != "completed":
+            raise ValueError(f"Batch {batch_id} is not completed yet (status: {batch.status})")
+
+        if not batch.output_file_id:
+            raise ValueError(f"Batch {batch_id} has no output file")
+
+        # Download results file
+        logger.debug(f"Downloading results for batch {batch_id} from file {batch.output_file_id}")
+        file_content = await self._client.files.content(batch.output_file_id)
+
+        # Parse JSONL results
+        results = []
+        for line in file_content.text.strip().split("\n"):
+            if line:
+                results.append(json.loads(line))
+
+        logger.info(f"Retrieved {len(results)} results for batch {batch_id}")
+
+        return results
+
+    async def cleanup(self) -> None:
+        """Clean up resources (close OpenAI client connections)."""
+        if hasattr(self, "_client") and self._client:
+            await self._client.close()

hindsight-api/hindsight_api/engine/query_analyzer.py

@@ -84,7 +84,7 @@ class DateparserQueryAnalyzer(QueryAnalyzer):
 
     Performance:
     - ~10-50ms per query
-    - No model loading required
+    - No model loading required (lazy import on first use)
     """
 
     def __init__(self):
@@ -112,8 +112,6 @@ class DateparserQueryAnalyzer(QueryAnalyzer):
         Returns:
             QueryAnalysis with temporal_constraint if found
         """
-        self.load()
-
         if reference_date is None:
             reference_date = datetime.now()
 
@@ -123,6 +121,9 @@ class DateparserQueryAnalyzer(QueryAnalyzer):
         if period_result is not None:
             return QueryAnalysis(temporal_constraint=period_result)
 
+        # Lazy load dateparser (only imports on first call, then cached)
+        self.load()
+
         # Use dateparser's search_dates to find temporal expressions
         settings = {
             "RELATIVE_BASE": reference_date,

hindsight-api/hindsight_api/engine/reflect/__init__.py

@@ -0,0 +1,18 @@
+"""
+Reflect agent module for agentic reflection with tools.
+
+The reflect agent uses an iterative loop with tools to:
+1. Lookup mental models (existing knowledge)
+2. Recall facts (semantic + temporal search)
+3. Expand memories (get chunk/document context)
+"""
+
+from .agent import ReflectAgentResult, run_reflect_agent
+from .models import ReflectAction, ReflectActionBatch
+
+__all__ = [
+    "run_reflect_agent",
+    "ReflectAgentResult",
+    "ReflectAction",
+    "ReflectActionBatch",
+]

hindsight-api/hindsight_api/engine/reflect/agent.py

@@ -0,0 +1,990 @@
+"""
+Reflect agent - agentic loop for reflection with native tool calling.
+
+Uses hierarchical retrieval:
+1. search_mental_models - User-curated summaries (highest quality)
+2. search_observations - Consolidated knowledge with freshness
+3. recall - Raw facts as ground truth
+"""
+
+import asyncio
+import json
+import logging
+import re
+import time
+from typing import TYPE_CHECKING, Any, Awaitable, Callable
+
+from .models import DirectiveInfo, LLMCall, ReflectAgentResult, TokenUsageSummary, ToolCall
+from .prompts import FINAL_SYSTEM_PROMPT, _extract_directive_rules, build_final_prompt, build_system_prompt_for_tools
+from .tools_schema import get_reflect_tools
+
+
+def _build_directives_applied(directives: list[dict[str, Any]] | None) -> list[DirectiveInfo]:
+    """Build list of DirectiveInfo from directive mental models.
+
+    Handles multiple directive formats:
+    1. New format: directives have direct 'content' field
+    2. Fallback: directives have 'description' field
+    """
+    if not directives:
+        return []
+
+    result = []
+    for directive in directives:
+        directive_id = directive.get("id", "")
+        directive_name = directive.get("name", "")
+
+        # Get content from 'content' field or fallback to 'description'
+        content = directive.get("content", "") or directive.get("description", "")
+
+        result.append(DirectiveInfo(id=directive_id, name=directive_name, content=content))
+
+    return result
+
+
+if TYPE_CHECKING:
+    from ..llm_wrapper import LLMProvider
+    from ..response_models import LLMToolCall
+
+logger = logging.getLogger(__name__)
+
+DEFAULT_MAX_ITERATIONS = 10
+
+
+def _normalize_tool_name(name: str) -> str:
+    """Normalize tool name from various LLM output formats.
+
+    Some LLMs output tool names in non-standard formats:
+    - 'functions.done' (OpenAI-style prefix)
+    - 'call=functions.done' (some models)
+    - 'call=done' (some models)
+    - 'done<|channel|>commentary' (malformed special tokens appended)
+
+    Returns the normalized tool name (e.g., 'done', 'recall', etc.)
+    """
+    # Handle 'call=functions.name' or 'call=name' format
+    if name.startswith("call="):
+        name = name[len("call=") :]
+
+    # Handle 'functions.name' format
+    if name.startswith("functions."):
+        name = name[len("functions.") :]
+
+    # Handle malformed special tokens appended to tool name
+    # e.g., 'done<|channel|>commentary' -> 'done'
+    if "<|" in name:
+        name = name.split("<|")[0]
+
+    return name
+
+
+def _is_done_tool(name: str) -> bool:
+    """Check if the tool name represents the 'done' tool."""
+    return _normalize_tool_name(name) == "done"
+
+
+# Pattern to match done() call as text - handles done({...}) with nested JSON
+_DONE_CALL_PATTERN = re.compile(r"done\s*\(\s*\{.*$", re.DOTALL)
+
+# Patterns for leaked structured output in the answer field
+_LEAKED_JSON_SUFFIX = re.compile(
+    r'\s*```(?:json)?\s*\{[^}]*(?:"(?:observation_ids|memory_ids|mental_model_ids)"|\})\s*```\s*$',
+    re.DOTALL | re.IGNORECASE,
+)
+_LEAKED_JSON_OBJECT = re.compile(
+    r'\s*\{[^{]*"(?:observation_ids|memory_ids|mental_model_ids|answer)"[^}]*\}\s*$', re.DOTALL
+)
+_TRAILING_IDS_PATTERN = re.compile(
+    r"\s*(?:observation_ids|memory_ids|mental_model_ids)\s*[=:]\s*\[.*?\]\s*$", re.DOTALL | re.IGNORECASE
+)
+
+
+def _clean_answer_text(text: str) -> str:
+    """Clean up answer text by removing any done() tool call syntax.
+
+    Some LLMs output the done() call as text instead of a proper tool call.
+    This strips out patterns like: done({"answer": "...", ...})
+    """
+    # Remove done() call pattern from the end of the text
+    cleaned = _DONE_CALL_PATTERN.sub("", text).strip()
+    return cleaned if cleaned else text
+
+
+def _clean_done_answer(text: str) -> str:
+    """Clean up the answer field from a done() tool call.
+
+    Some LLMs leak structured output patterns into the answer text, such as:
+    - JSON code blocks with observation_ids/memory_ids at the end
+    - Raw JSON objects with these fields
+    - Plain text like "observation_ids: [...]"
+
+    This cleans those patterns while preserving the actual answer content.
+    """
+    if not text:
+        return text
+
+    cleaned = text
+
+    # Remove leaked JSON in code blocks at the end
+    cleaned = _LEAKED_JSON_SUFFIX.sub("", cleaned).strip()
+
+    # Remove leaked raw JSON objects at the end
+    cleaned = _LEAKED_JSON_OBJECT.sub("", cleaned).strip()
+
+    # Remove trailing ID patterns
+    cleaned = _TRAILING_IDS_PATTERN.sub("", cleaned).strip()
+
+    return cleaned if cleaned else text
+
+
+async def _generate_structured_output(
+    answer: str,
+    response_schema: dict,
+    llm_config: "LLMProvider",
+    reflect_id: str,
+) -> tuple[dict[str, Any] | None, int, int]:
+    """Generate structured output from an answer using the provided JSON schema.
+
+    Args:
+        answer: The text answer to extract structured data from
+        response_schema: JSON Schema for the expected output structure
+        llm_config: LLM provider for making the extraction call
+        reflect_id: Reflect ID for logging
+
+    Returns:
+        Tuple of (structured_output, input_tokens, output_tokens).
+        structured_output is None if generation fails.
+    """
+    try:
+        from typing import Any as TypingAny
+
+        from pydantic import create_model
+
+        def _json_schema_type_to_python(field_schema: dict) -> type:
+            """Map JSON schema type to Python type for better LLM guidance."""
+            json_type = field_schema.get("type", "string")
+            if json_type == "array":
+                return list
+            elif json_type == "object":
+                return dict
+            elif json_type == "integer":
+                return int
+            elif json_type == "number":
+                return float
+            elif json_type == "boolean":
+                return bool
+            else:
+                return str
+
+        # Build fields from JSON schema properties
+        schema_props = response_schema.get("properties", {})
+        required_fields = set(response_schema.get("required", []))
+        fields: dict[str, TypingAny] = {}
+        for field_name, field_schema in schema_props.items():
+            field_type = _json_schema_type_to_python(field_schema)
+            default = ... if field_name in required_fields else None
+            fields[field_name] = (field_type, default)
+
+        if not fields:
+            logger.warning(f"[REFLECT {reflect_id}] No fields found in response_schema, skipping structured output")
+            return None, 0, 0
+
+        DynamicModel = create_model("StructuredResponse", **fields)
+
+        # Include the full schema in the prompt for better LLM guidance
+        schema_str = json.dumps(response_schema, indent=2)
+
+        # Build field descriptions for the prompt
+        field_descriptions = []
+        for field_name, field_schema in schema_props.items():
+            field_type = field_schema.get("type", "string")
+            field_desc = field_schema.get("description", "")
+            is_required = field_name in required_fields
+            req_marker = " (REQUIRED)" if is_required else " (optional)"
+            field_descriptions.append(f"- {field_name} ({field_type}){req_marker}: {field_desc}")
+        fields_text = "\n".join(field_descriptions)
+
+        # Call LLM with the answer to extract structured data
+        structured_prompt = f"""Your task is to extract specific information from the answer below and format it as JSON.
+
+ANSWER TO EXTRACT FROM:
+\"\"\"
+{answer}
+\"\"\"
+
+REQUIRED OUTPUT FORMAT - Extract the following fields from the answer above:
+{fields_text}
+
+JSON Schema:
+```json
+{schema_str}
+```
+
+INSTRUCTIONS:
+1. Read the answer carefully and identify the information that matches each field
+2. Extract the ACTUAL content from the answer - do NOT leave fields empty if information is present
+3. For string fields: use the exact text or a clear summary from the answer
+4. For array fields: return a JSON array (e.g., ["item1", "item2"]), NOT a string
+5. For required fields: you MUST provide a value extracted from the answer
+6. Return ONLY the JSON object, no explanation
+
+OUTPUT:"""
+
+        structured_result, usage = await llm_config.call(
+            messages=[
+                {
+                    "role": "system",
+                    "content": "You are a precise data extraction assistant. Extract information from text and return it as valid JSON matching the provided schema. Always extract actual content - never return empty strings for required fields if information is available.",
+                },
+                {"role": "user", "content": structured_prompt},
+            ],
+            response_format=DynamicModel,
+            scope="reflect_structured",
+            skip_validation=True,  # We'll handle the dict ourselves
+            return_usage=True,
+        )
+
+        # Convert to dict
+        if hasattr(structured_result, "model_dump"):
+            structured_output = structured_result.model_dump()
+        elif isinstance(structured_result, dict):
+            structured_output = structured_result
+        else:
+            # Try to parse as JSON
+            structured_output = json.loads(str(structured_result))
+
+        # Validate that required fields have non-empty values
+        for field_name in required_fields:
+            value = structured_output.get(field_name)
+            if value is None or value == "" or value == []:
+                logger.warning(f"[REFLECT {reflect_id}] Required field '{field_name}' is empty in structured output")
+
+        logger.info(f"[REFLECT {reflect_id}] Generated structured output with {len(structured_output)} fields")
+        return structured_output, usage.input_tokens, usage.output_tokens
+
+    except Exception as e:
+        logger.warning(f"[REFLECT {reflect_id}] Failed to generate structured output: {e}")
+        return None, 0, 0
+
+
+async def run_reflect_agent(
+    llm_config: "LLMProvider",
+    bank_id: str,
+    query: str,
+    bank_profile: dict[str, Any],
+    search_mental_models_fn: Callable[[str, int], Awaitable[dict[str, Any]]],
+    search_observations_fn: Callable[[str, int], Awaitable[dict[str, Any]]],
+    recall_fn: Callable[[str, int], Awaitable[dict[str, Any]]],
+    expand_fn: Callable[[list[str], str], Awaitable[dict[str, Any]]],
+    context: str | None = None,
+    max_iterations: int = DEFAULT_MAX_ITERATIONS,
+    max_tokens: int | None = None,
+    response_schema: dict | None = None,
+    directives: list[dict[str, Any]] | None = None,
+    has_mental_models: bool = False,
+    budget: str | None = None,
+) -> ReflectAgentResult:
+    """
+    Execute the reflect agent loop using native tool calling.
+
+    The agent uses hierarchical retrieval:
+    1. search_mental_models - User-curated summaries (try first)
+    2. search_observations - Consolidated knowledge with freshness
+    3. recall - Raw facts as ground truth
+
+    Args:
+        llm_config: LLM provider for agent calls
+        bank_id: Bank identifier
+        query: Question to answer
+        bank_profile: Bank profile with name and mission
+        search_mental_models_fn: Tool callback for searching mental models (query, max_results) -> result
+        search_observations_fn: Tool callback for searching observations (query, max_results) -> result
+        recall_fn: Tool callback for recall (query, max_tokens) -> result
+        expand_fn: Tool callback for expand (memory_ids, depth) -> result
+        context: Optional additional context
+        max_iterations: Maximum number of iterations before forcing response
+        max_tokens: Maximum tokens for the final response
+        response_schema: Optional JSON Schema for structured output in final response
+        directives: Optional list of directive mental models to inject as hard rules
+
+    Returns:
+        ReflectAgentResult with final answer and metadata
+    """
+    reflect_id = f"{bank_id[:8]}-{int(time.time() * 1000) % 100000}"
+    start_time = time.time()
+
+    # Build directives_applied for the trace
+    directives_applied = _build_directives_applied(directives)
+
+    # Extract directive rules for tool schema (if any)
+    directive_rules = _extract_directive_rules(directives) if directives else None
+
+    # Get tools for this agent (with directive compliance field if directives exist)
+    tools = get_reflect_tools(directive_rules=directive_rules)
+
+    # Build initial messages (directives are injected into system prompt at START and END)
+    system_prompt = build_system_prompt_for_tools(
+        bank_profile, context, directives=directives, has_mental_models=has_mental_models, budget=budget
+    )
+    messages: list[dict[str, Any]] = [
+        {"role": "system", "content": system_prompt},
+        {"role": "user", "content": query},
+    ]
+
+    # Tracking
+    total_tools_called = 0
+    tool_trace: list[ToolCall] = []
+    tool_trace_summary: list[dict[str, Any]] = []
+    llm_trace: list[dict[str, Any]] = []
+    context_history: list[dict[str, Any]] = []  # For final prompt fallback
+
+    # Token usage tracking - accumulate across all LLM calls
+    total_input_tokens = 0
+    total_output_tokens = 0
+
+    # Track available IDs for validation (prevents hallucinated citations)
+    available_memory_ids: set[str] = set()
+    available_mental_model_ids: set[str] = set()
+    available_observation_ids: set[str] = set()
+
+    def _get_llm_trace() -> list[LLMCall]:
+        return [
+            LLMCall(
+                scope=c["scope"],
+                duration_ms=c["duration_ms"],
+                input_tokens=c.get("input_tokens", 0),
+                output_tokens=c.get("output_tokens", 0),
+            )
+            for c in llm_trace
+        ]
+
+    def _get_usage() -> TokenUsageSummary:
+        return TokenUsageSummary(
+            input_tokens=total_input_tokens,
+            output_tokens=total_output_tokens,
+            total_tokens=total_input_tokens + total_output_tokens,
+        )
+
+    def _log_completion(answer: str, iterations: int, forced: bool = False):
+        elapsed_ms = int((time.time() - start_time) * 1000)
+        tools_summary = (
+            ", ".join(
+                f"{t['tool']}({t['input_summary']})={t['duration_ms']}ms/{t.get('output_chars', 0)}c"
+                for t in tool_trace_summary
+            )
+            or "none"
+        )
+        llm_summary = ", ".join(f"{c['scope']}={c['duration_ms']}ms" for c in llm_trace) or "none"
+        total_llm_ms = sum(c["duration_ms"] for c in llm_trace)
+        total_tools_ms = sum(t["duration_ms"] for t in tool_trace_summary)
+
+        answer_preview = answer[:100] + "..." if len(answer) > 100 else answer
+        mode = "forced" if forced else "done"
+        logger.info(
+            f"[REFLECT {reflect_id}] {mode} | "
+            f"query='{query[:50]}...' | "
+            f"iterations={iterations} | "
+            f"llm=[{llm_summary}] ({total_llm_ms}ms) | "
+            f"tools=[{tools_summary}] ({total_tools_ms}ms) | "
+            f"answer='{answer_preview}' | "
+            f"total={elapsed_ms}ms"
+        )
+
+    for iteration in range(max_iterations):
+        is_last = iteration == max_iterations - 1
+
+        if is_last:
+            # Force text response on last iteration - no tools
+            prompt = build_final_prompt(query, context_history, bank_profile, context)
+            llm_start = time.time()
+            response, usage = await llm_config.call(
+                messages=[
+                    {"role": "system", "content": FINAL_SYSTEM_PROMPT},
+                    {"role": "user", "content": prompt},
+                ],
+                scope="reflect",
+                max_completion_tokens=max_tokens,
+                return_usage=True,
+            )
+            llm_duration = int((time.time() - llm_start) * 1000)
+            total_input_tokens += usage.input_tokens
+            total_output_tokens += usage.output_tokens
+            llm_trace.append(
+                {
+                    "scope": "final",
+                    "duration_ms": llm_duration,
+                    "input_tokens": usage.input_tokens,
+                    "output_tokens": usage.output_tokens,
+                }
+            )
+            answer = _clean_answer_text(response.strip())
+
+            # Generate structured output if schema provided
+            structured_output = None
+            if response_schema and answer:
+                structured_output, struct_in, struct_out = await _generate_structured_output(
+                    answer, response_schema, llm_config, reflect_id
+                )
+                total_input_tokens += struct_in
+                total_output_tokens += struct_out
+
+            _log_completion(answer, iteration + 1, forced=True)
+            return ReflectAgentResult(
+                text=answer,
+                structured_output=structured_output,
+                iterations=iteration + 1,
+                tools_called=total_tools_called,
+                tool_trace=tool_trace,
+                llm_trace=_get_llm_trace(),
+                usage=_get_usage(),
+                directives_applied=directives_applied,
+            )
+
+        # Call LLM with tools
+        llm_start = time.time()
+
+        try:
+            result = await llm_config.call_with_tools(
+                messages=messages,
+                tools=tools,
+                scope="reflect_tool_call",
+                tool_choice="required" if iteration == 0 else "auto",  # Force tool use on first iteration
+            )
+            llm_duration = int((time.time() - llm_start) * 1000)
+            total_input_tokens += result.input_tokens
+            total_output_tokens += result.output_tokens
+            llm_trace.append(
+                {
+                    "scope": f"agent_{iteration + 1}",
+                    "duration_ms": llm_duration,
+                    "input_tokens": result.input_tokens,
+                    "output_tokens": result.output_tokens,
+                }
+            )
+
+        except Exception as e:
+            err_duration = int((time.time() - llm_start) * 1000)
+            logger.warning(f"[REFLECT {reflect_id}] LLM error on iteration {iteration + 1}: {e} ({err_duration}ms)")
+            llm_trace.append({"scope": f"agent_{iteration + 1}_err", "duration_ms": err_duration})
+            # Guardrail: If no evidence gathered yet, retry
+            has_gathered_evidence = (
+                bool(available_memory_ids) or bool(available_mental_model_ids) or bool(available_observation_ids)
+            )
+            if not has_gathered_evidence and iteration < max_iterations - 1:
+                continue
+            prompt = build_final_prompt(query, context_history, bank_profile, context)
+            llm_start = time.time()
+            response, usage = await llm_config.call(
+                messages=[
+                    {"role": "system", "content": FINAL_SYSTEM_PROMPT},
+                    {"role": "user", "content": prompt},
+                ],
+                scope="reflect",
+                max_completion_tokens=max_tokens,
+                return_usage=True,
+            )
+            llm_duration = int((time.time() - llm_start) * 1000)
+            total_input_tokens += usage.input_tokens
+            total_output_tokens += usage.output_tokens
+            llm_trace.append(
+                {
+                    "scope": "final",
+                    "duration_ms": llm_duration,
+                    "input_tokens": usage.input_tokens,
+                    "output_tokens": usage.output_tokens,
+                }
+            )
+            answer = _clean_answer_text(response.strip())
+
+            # Generate structured output if schema provided
+            structured_output = None
+            if response_schema and answer:
+                structured_output, struct_in, struct_out = await _generate_structured_output(
+                    answer, response_schema, llm_config, reflect_id
+                )
+                total_input_tokens += struct_in
+                total_output_tokens += struct_out
+
+            _log_completion(answer, iteration + 1, forced=True)
+            return ReflectAgentResult(
+                text=answer,
+                structured_output=structured_output,
+                iterations=iteration + 1,
+                tools_called=total_tools_called,
+                tool_trace=tool_trace,
+                llm_trace=_get_llm_trace(),
+                usage=_get_usage(),
+                directives_applied=directives_applied,
+            )
+
+        # No tool calls - LLM wants to respond with text
+        if not result.tool_calls:
+            if result.content:
+                answer = _clean_answer_text(result.content.strip())
+
+                # Generate structured output if schema provided
+                structured_output = None
+                if response_schema and answer:
+                    structured_output, struct_in, struct_out = await _generate_structured_output(
+                        answer, response_schema, llm_config, reflect_id
+                    )
+                    total_input_tokens += struct_in
+                    total_output_tokens += struct_out
+
+                _log_completion(answer, iteration + 1)
+                return ReflectAgentResult(
+                    text=answer,
+                    structured_output=structured_output,
+                    iterations=iteration + 1,
+                    tools_called=total_tools_called,
+                    tool_trace=tool_trace,
+                    llm_trace=_get_llm_trace(),
+                    usage=_get_usage(),
+                    directives_applied=directives_applied,
+                )
+            # Empty response, force final
+            prompt = build_final_prompt(query, context_history, bank_profile, context)
+            llm_start = time.time()
+            response, usage = await llm_config.call(
+                messages=[
+                    {"role": "system", "content": FINAL_SYSTEM_PROMPT},
+                    {"role": "user", "content": prompt},
+                ],
+                scope="reflect",
+                max_completion_tokens=max_tokens,
+                return_usage=True,
+            )
+            llm_duration = int((time.time() - llm_start) * 1000)
+            total_input_tokens += usage.input_tokens
+            total_output_tokens += usage.output_tokens
+            llm_trace.append(
+                {
+                    "scope": "final",
+                    "duration_ms": llm_duration,
+                    "input_tokens": usage.input_tokens,
+                    "output_tokens": usage.output_tokens,
+                }
+            )
+            answer = _clean_answer_text(response.strip())
+
+            # Generate structured output if schema provided
+            structured_output = None
+            if response_schema and answer:
+                structured_output, struct_in, struct_out = await _generate_structured_output(
+                    answer, response_schema, llm_config, reflect_id
+                )
+                total_input_tokens += struct_in
+                total_output_tokens += struct_out
+
+            _log_completion(answer, iteration + 1, forced=True)
+            return ReflectAgentResult(
+                text=answer,
+                structured_output=structured_output,
+                iterations=iteration + 1,
+                tools_called=total_tools_called,
+                tool_trace=tool_trace,
+                llm_trace=_get_llm_trace(),
+                usage=_get_usage(),
+                directives_applied=directives_applied,
+            )
+
+        # Check for done tool call (handle various LLM output formats)
+        done_call = next((tc for tc in result.tool_calls if _is_done_tool(tc.name)), None)
+        if done_call:
+            # Guardrail: Require evidence before done
+            has_gathered_evidence = (
+                bool(available_memory_ids) or bool(available_mental_model_ids) or bool(available_observation_ids)
+            )
+            if not has_gathered_evidence and iteration < max_iterations - 1:
+                # Add assistant message and fake tool result asking for evidence
+                messages.append(
+                    {
+                        "role": "assistant",
+                        "tool_calls": [_tool_call_to_dict(done_call)],
+                    }
+                )
+                messages.append(
+                    {
+                        "role": "tool",
+                        "tool_call_id": done_call.id,
+                        "name": done_call.name,  # Required by Gemini
+                        "content": json.dumps(
+                            {
+                                "error": "You must search for information first. Use search_mental_models(), search_observations(), or recall() before providing your final answer."
+                            }
+                        ),
+                    }
+                )
+                continue
+
+            # Process done tool - wrap with tool call span
+            from hindsight_api.tracing import get_tracer
+
+            tracer = get_tracer()
+            span_name = "hindsight.reflect_tool_call"
+            with tracer.start_as_current_span(span_name) as span:
+                span.set_attribute("hindsight.scope", "reflect_tool_call")
+                span.set_attribute("hindsight.operation", "reflect_tool_call")
+                return await _process_done_tool(
+                    done_call,
+                    available_memory_ids,
+                    available_mental_model_ids,
+                    available_observation_ids,
+                    iteration + 1,
+                    total_tools_called,
+                    tool_trace,
+                    _get_llm_trace(),
+                    _get_usage(),
+                    _log_completion,
+                    reflect_id,
+                    directives_applied=directives_applied,
+                    llm_config=llm_config,
+                    response_schema=response_schema,
+                )
+
+        # Execute other tools in parallel (exclude done tool in all its format variants)
+        other_tools = [tc for tc in result.tool_calls if not _is_done_tool(tc.name)]
+        if other_tools:
+            # Add assistant message with tool calls
+            messages.append(
+                {
+                    "role": "assistant",
+                    "tool_calls": [_tool_call_to_dict(tc) for tc in other_tools],
+                }
+            )
+
+            # Execute tools in parallel
+            tool_tasks = [
+                _execute_tool_with_timing(
+                    tc,
+                    search_mental_models_fn,
+                    search_observations_fn,
+                    recall_fn,
+                    expand_fn,
+                )
+                for tc in other_tools
+            ]
+            tool_results = await asyncio.gather(*tool_tasks, return_exceptions=True)
+            total_tools_called += len(other_tools)
+
+            # Process results and add to messages
+            for tc, result_data in zip(other_tools, tool_results):
+                if isinstance(result_data, Exception):
+                    # Tool execution failed - send error back to LLM so it can try again
+                    logger.warning(f"[REFLECT {reflect_id}] Tool {tc.name} failed with exception: {result_data}")
+                    output = {"error": f"Tool execution failed: {result_data}"}
+                    duration_ms = 0
+                else:
+                    output, duration_ms = result_data
+
+                # Normalize tool name for consistent tracking
+                normalized_tool_name = _normalize_tool_name(tc.name)
+
+                # Check if tool returned an error response - log but continue (LLM will see the error)
+                if isinstance(output, dict) and "error" in output:
+                    logger.warning(
+                        f"[REFLECT {reflect_id}] Tool {normalized_tool_name} returned error: {output['error']}"
+                    )
+
+                # Track available IDs from tool results (only for successful responses)
+                if (
+                    normalized_tool_name == "search_mental_models"
+                    and isinstance(output, dict)
+                    and "mental_models" in output
+                ):
+                    for mm in output["mental_models"]:
+                        if "id" in mm:
+                            available_mental_model_ids.add(mm["id"])
+
+                if (
+                    normalized_tool_name == "search_observations"
+                    and isinstance(output, dict)
+                    and "observations" in output
+                ):
+                    for obs in output["observations"]:
+                        if "id" in obs:
+                            available_observation_ids.add(obs["id"])
+
+                if normalized_tool_name == "recall" and isinstance(output, dict) and "memories" in output:
+                    for memory in output["memories"]:
+                        if "id" in memory:
+                            available_memory_ids.add(memory["id"])
+
+                # Add tool result message
+                messages.append(
+                    {
+                        "role": "tool",
+                        "tool_call_id": tc.id,
+                        "name": tc.name,  # Required by Gemini
+                        "content": json.dumps(output, default=str),
+                    }
+                )
+
+                # Track for logging and context history
+                input_dict = {"tool": tc.name, **tc.arguments}
+                input_summary = _summarize_input(tc.name, tc.arguments)
+
+                # Extract reason from tool arguments (if provided)
+                tool_reason = tc.arguments.get("reason")
+
+                tool_trace.append(
+                    ToolCall(
+                        tool=tc.name,
+                        reason=tool_reason,
+                        input=input_dict,
+                        output=output,
+                        duration_ms=duration_ms,
+                        iteration=iteration + 1,
+                    )
+                )
+
+                try:
+                    output_chars = len(json.dumps(output))
+                except (TypeError, ValueError):
+                    output_chars = len(str(output))
+
+                tool_trace_summary.append(
+                    {
+                        "tool": tc.name,
+                        "input_summary": input_summary,
+                        "duration_ms": duration_ms,
+                        "output_chars": output_chars,
+                    }
+                )
+
+                # Keep context history for fallback final prompt
+                context_history.append({"tool": tc.name, "input": input_dict, "output": output})
+
+    # Should not reach here
+    answer = "I was unable to formulate a complete answer within the iteration limit."
+    _log_completion(answer, max_iterations, forced=True)
+    return ReflectAgentResult(
+        text=answer,
+        iterations=max_iterations,
+        tools_called=total_tools_called,
+        tool_trace=tool_trace,
+        llm_trace=_get_llm_trace(),
+        usage=_get_usage(),
+        directives_applied=directives_applied,
+    )
+
+
+def _tool_call_to_dict(tc: "LLMToolCall") -> dict[str, Any]:
+    """Convert LLMToolCall to OpenAI message format."""
+    return {
+        "id": tc.id,
+        "type": "function",
+        "function": {
+            "name": tc.name,
+            "arguments": json.dumps(tc.arguments),
+        },
+    }
+
+
+async def _process_done_tool(
+    done_call: "LLMToolCall",
+    available_memory_ids: set[str],
+    available_mental_model_ids: set[str],
+    available_observation_ids: set[str],
+    iterations: int,
+    total_tools_called: int,
+    tool_trace: list[ToolCall],
+    llm_trace: list[LLMCall],
+    usage: TokenUsageSummary,
+    log_completion: Callable,
+    reflect_id: str,
+    directives_applied: list[DirectiveInfo],
+    llm_config: "LLMProvider | None" = None,
+    response_schema: dict | None = None,
+) -> ReflectAgentResult:
+    """Process the done tool call and return the result."""
+    args = done_call.arguments
+
+    # Extract and clean the answer - some LLMs leak structured output into the answer text
+    raw_answer = args.get("answer", "").strip()
+    answer = _clean_done_answer(raw_answer) if raw_answer else ""
+    if not answer:
+        answer = "No answer provided."
+
+    # Validate IDs (only include IDs that were actually retrieved)
+    used_memory_ids = [mid for mid in args.get("memory_ids", []) if mid in available_memory_ids]
+    used_mental_model_ids = [mid for mid in args.get("mental_model_ids", []) if mid in available_mental_model_ids]
+    used_observation_ids = [oid for oid in args.get("observation_ids", []) if oid in available_observation_ids]
+
+    # Generate structured output if schema provided
+    structured_output = None
+    final_usage = usage
+    if response_schema and llm_config and answer:
+        structured_output, struct_in, struct_out = await _generate_structured_output(
+            answer, response_schema, llm_config, reflect_id
+        )
+        # Add structured output tokens to usage
+        final_usage = TokenUsageSummary(
+            input_tokens=usage.input_tokens + struct_in,
+            output_tokens=usage.output_tokens + struct_out,
+            total_tokens=usage.total_tokens + struct_in + struct_out,
+        )
+
+    log_completion(answer, iterations)
+    return ReflectAgentResult(
+        text=answer,
+        structured_output=structured_output,
+        iterations=iterations,
+        tools_called=total_tools_called,
+        tool_trace=tool_trace,
+        llm_trace=llm_trace,
+        usage=final_usage,
+        used_memory_ids=used_memory_ids,
+        used_mental_model_ids=used_mental_model_ids,
+        used_observation_ids=used_observation_ids,
+        directives_applied=directives_applied,
+    )
+
+
+async def _execute_tool_with_timing(
+    tc: "LLMToolCall",
+    search_mental_models_fn: Callable[[str, int], Awaitable[dict[str, Any]]],
+    search_observations_fn: Callable[[str, int], Awaitable[dict[str, Any]]],
+    recall_fn: Callable[[str, int], Awaitable[dict[str, Any]]],
+    expand_fn: Callable[[list[str], str], Awaitable[dict[str, Any]]],
+) -> tuple[dict[str, Any], int]:
+    """Execute a tool call and return result with timing."""
+    from hindsight_api.tracing import get_tracer
+
+    start_time = time.time()
+
+    # Create span for tool execution
+    tracer = get_tracer()
+    # Normalize tool name for span
+    normalized_name = _normalize_tool_name(tc.name)
+    span_name = f"hindsight.reflect_tool_exec.{normalized_name}"
+
+    # Calculate timestamps
+    start_time_ns = time.time_ns()
+
+    with tracer.start_as_current_span(
+        span_name,
+        start_time=start_time_ns,
+        end_on_exit=False,
+    ) as span:
+        # Set attributes
+        span.set_attribute("hindsight.tool.name", normalized_name)
+        span.set_attribute("hindsight.tool.id", tc.id)
+        span.set_attribute("hindsight.tool.arguments", json.dumps(tc.arguments))
+
+        try:
+            result = await _execute_tool(
+                tc.name,
+                tc.arguments,
+                search_mental_models_fn,
+                search_observations_fn,
+                recall_fn,
+                expand_fn,
+            )
+
+            # Set success attributes
+            if isinstance(result, dict) and "error" in result:
+                from opentelemetry.trace import Status, StatusCode
+
+                span.set_status(Status(StatusCode.ERROR, result["error"]))
+            else:
+                from opentelemetry.trace import Status, StatusCode
+
+                span.set_status(Status(StatusCode.OK))
+
+            duration_ms = int((time.time() - start_time) * 1000)
+            span.set_attribute("hindsight.tool.duration_ms", duration_ms)
+
+            # End span with correct timestamp
+            end_time_ns = time.time_ns()
+            span.end(end_time=end_time_ns)
+
+            return result, duration_ms
+        except Exception as e:
+            from opentelemetry.trace import Status, StatusCode
+
+            span.set_status(Status(StatusCode.ERROR, str(e)))
+            span.record_exception(e)
+            duration_ms = int((time.time() - start_time) * 1000)
+            span.set_attribute("hindsight.tool.duration_ms", duration_ms)
+            end_time_ns = time.time_ns()
+            span.end(end_time=end_time_ns)
+            raise
+
+
+async def _execute_tool(
+    tool_name: str,
+    args: dict[str, Any],
+    search_mental_models_fn: Callable[[str, int], Awaitable[dict[str, Any]]],
+    search_observations_fn: Callable[[str, int], Awaitable[dict[str, Any]]],
+    recall_fn: Callable[[str, int], Awaitable[dict[str, Any]]],
+    expand_fn: Callable[[list[str], str], Awaitable[dict[str, Any]]],
+) -> dict[str, Any]:
+    """Execute a single tool by name."""
+    # Normalize tool name for various LLM output formats
+    tool_name = _normalize_tool_name(tool_name)
+
+    if tool_name == "search_mental_models":
+        query = args.get("query")
+        if not query:
+            return {"error": "search_mental_models requires a query parameter"}
+        max_results = int(args.get("max_results") or 5)
+        return await search_mental_models_fn(query, max_results)
+
+    elif tool_name == "search_observations":
+        query = args.get("query")
+        if not query:
+            return {"error": "search_observations requires a query parameter"}
+        max_tokens = max(int(args.get("max_tokens") or 5000), 1000)  # Default 5000, min 1000
+        return await search_observations_fn(query, max_tokens)
+
+    elif tool_name == "recall":
+        query = args.get("query")
+        if not query:
+            return {"error": "recall requires a query parameter"}
+        max_tokens = max(int(args.get("max_tokens") or 2048), 1000)  # Default 2048, min 1000
+        return await recall_fn(query, max_tokens)
+
+    elif tool_name == "expand":
+        memory_ids = args.get("memory_ids", [])
+        if not memory_ids:
+            return {"error": "expand requires memory_ids"}
+        depth = args.get("depth", "chunk")
+        return await expand_fn(memory_ids, depth)
+
+    else:
+        return {"error": f"Unknown tool: {tool_name}"}
+
+
+def _summarize_input(tool_name: str, args: dict[str, Any]) -> str:
+    """Create a summary of tool input for logging, showing all params."""
+    if tool_name == "search_mental_models":
+        query = args.get("query", "")
+        query_preview = f"'{query[:30]}...'" if len(query) > 30 else f"'{query}'"
+        max_results = int(args.get("max_results") or 5)
+        return f"(query={query_preview}, max_results={max_results})"
+    elif tool_name == "search_observations":
+        query = args.get("query", "")
+        query_preview = f"'{query[:30]}...'" if len(query) > 30 else f"'{query}'"
+        max_tokens = max(int(args.get("max_tokens") or 5000), 1000)
+        return f"(query={query_preview}, max_tokens={max_tokens})"
+    elif tool_name == "recall":
+        query = args.get("query", "")
+        query_preview = f"'{query[:30]}...'" if len(query) > 30 else f"'{query}'"
+        # Show actual value used (default 2048, min 1000)
+        max_tokens = max(int(args.get("max_tokens") or 2048), 1000)
+        return f"(query={query_preview}, max_tokens={max_tokens})"
+    elif tool_name == "expand":
+        memory_ids = args.get("memory_ids", [])
+        depth = args.get("depth", "chunk")
+        return f"(memory_ids=[{len(memory_ids)} ids], depth={depth})"
+    elif tool_name == "done":
+        answer = args.get("answer", "")
+        answer_preview = f"'{answer[:30]}...'" if len(answer) > 30 else f"'{answer}'"
+        memory_ids = args.get("memory_ids", [])
+        mental_model_ids = args.get("mental_model_ids", [])
+        observation_ids = args.get("observation_ids", [])
+        return (
+            f"(answer={answer_preview}, mem={len(memory_ids)}, mm={len(mental_model_ids)}, obs={len(observation_ids)})"
+        )
+    return str(args)

hindsight-api/hindsight_api/engine/reflect/models.py

@@ -0,0 +1,109 @@
+"""
+Pydantic models for the reflect agent.
+"""
+
+from typing import Any, Literal
+
+from pydantic import BaseModel, Field
+
+
+class ObservationSection(BaseModel):
+    """A section within an observation with its supporting memories."""
+
+    title: str = Field(description="Section header (can be empty for intro)")
+    text: str = Field(description="Section content - no headers, use lists/tables/bold")
+    memory_ids: list[str] = Field(default_factory=list, description="Memory IDs supporting this section")
+
+
+class ReflectAction(BaseModel):
+    """Single action the reflect agent can take."""
+
+    tool: Literal["list_observations", "get_observation", "recall", "expand", "done"] = Field(
+        description="Tool to invoke: list_observations, get_observation, recall, expand, or done"
+    )
+    # Tool-specific parameters
+    observation_id: str | None = Field(default=None, description="Observation ID for get_observation")
+    query: str | None = Field(default=None, description="Search query for recall")
+    max_tokens: int | None = Field(default=None, description="Max tokens for recall results (default 2048)")
+    memory_ids: list[str] | None = Field(default=None, description="Memory unit IDs for expand (batched)")
+    depth: Literal["chunk", "document"] | None = Field(default=None, description="Expansion depth for expand")
+    observation_sections: list[ObservationSection] | None = Field(
+        default=None, description="Observation sections for done action (when output_mode=observations)"
+    )
+    # Plain text answer fields (for output_mode=answer)
+    answer: str | None = Field(default=None, description="Well-formatted markdown answer for done action")
+    answer_memory_ids: list[str] | None = Field(
+        default=None, description="Memory IDs supporting the answer", alias="memory_ids"
+    )
+    answer_model_ids: list[str] | None = Field(
+        default=None, description="Mental model IDs supporting the answer", alias="model_ids"
+    )
+    reasoning: str | None = Field(default=None, description="Brief reasoning for this action")
+
+
+class ReflectActionBatch(BaseModel):
+    """Batch of actions for parallel execution."""
+
+    actions: list[ReflectAction] = Field(description="List of actions to execute in parallel")
+
+
+class ToolCall(BaseModel):
+    """A single tool call made during reflect."""
+
+    tool: str = Field(description="Tool name: lookup, recall, expand")
+    reason: str | None = Field(default=None, description="Agent's reasoning for making this tool call")
+    input: dict = Field(description="Tool input parameters")
+    output: dict = Field(description="Tool output/result")
+    duration_ms: int = Field(description="Execution time in milliseconds")
+    iteration: int = Field(default=0, description="Iteration number (1-based) when this tool was called")
+
+
+class LLMCall(BaseModel):
+    """A single LLM call made during reflect."""
+
+    scope: str = Field(description="Call scope: agent_1, agent_2, final, etc.")
+    duration_ms: int = Field(description="Execution time in milliseconds")
+    input_tokens: int = Field(default=0, description="Input tokens used")
+    output_tokens: int = Field(default=0, description="Output tokens used")
+
+
+class DirectiveInfo(BaseModel):
+    """Information about a directive that was applied during reflect."""
+
+    id: str = Field(description="Directive mental model ID")
+    name: str = Field(description="Directive name")
+    content: str = Field(description="Directive content")
+
+
+class TokenUsageSummary(BaseModel):
+    """Total token usage across all LLM calls."""
+
+    input_tokens: int = Field(default=0, description="Total input tokens used")
+    output_tokens: int = Field(default=0, description="Total output tokens used")
+    total_tokens: int = Field(default=0, description="Total tokens (input + output)")
+
+
+class ReflectAgentResult(BaseModel):
+    """Result from the reflect agent."""
+
+    text: str = Field(description="Final answer text")
+    structured_output: dict[str, Any] | None = Field(
+        default=None, description="Structured output parsed according to provided response_schema"
+    )
+    iterations: int = Field(default=0, description="Number of iterations taken")
+    tools_called: int = Field(default=0, description="Total number of tool calls made")
+    tool_trace: list[ToolCall] = Field(default_factory=list, description="Trace of all tool calls made")
+    llm_trace: list[LLMCall] = Field(default_factory=list, description="Trace of all LLM calls made")
+    usage: TokenUsageSummary = Field(
+        default_factory=TokenUsageSummary, description="Total token usage across all LLM calls"
+    )
+    used_memory_ids: list[str] = Field(default_factory=list, description="Validated memory IDs actually used in answer")
+    used_mental_model_ids: list[str] = Field(
+        default_factory=list, description="Validated mental model IDs actually used in answer"
+    )
+    used_observation_ids: list[str] = Field(
+        default_factory=list, description="Validated observation IDs actually used in answer"
+    )
+    directives_applied: list[DirectiveInfo] = Field(
+        default_factory=list, description="Directive mental models that affected this reflection"
+    )

hindsight-api/hindsight_api/engine/reflect/observations.py

@@ -0,0 +1,186 @@
+"""
+Models and utilities for evidence-grounded observations with computed trends.
+
+Observations are part of mental models and represent patterns/beliefs derived
+from memories. Each observation must be grounded in specific evidence (quotes)
+from memories, and trends are computed algorithmically from evidence timestamps.
+"""
+
+from datetime import datetime, timedelta, timezone
+from enum import Enum
+
+from pydantic import BaseModel, Field, computed_field, field_validator
+
+
+class Trend(str, Enum):
+    """Computed trend for an observation based on evidence timestamps.
+
+    Trends indicate how an observation's evidence is distributed over time:
+    - STABLE: Evidence spread across time, continues to present
+    - STRENGTHENING: More/denser evidence recently than before
+    - WEAKENING: Evidence mostly old, sparse recently
+    - NEW: All evidence within recent window
+    - STALE: No evidence in recent window (may no longer apply)
+    """
+
+    STABLE = "stable"
+    STRENGTHENING = "strengthening"
+    WEAKENING = "weakening"
+    NEW = "new"
+    STALE = "stale"
+
+
+class ObservationEvidence(BaseModel):
+    """A single piece of evidence supporting an observation.
+
+    Each evidence item must include an exact quote from the source memory
+    to ensure observations are grounded and verifiable.
+    """
+
+    memory_id: str = Field(description="ID of the memory unit this evidence comes from")
+    quote: str = Field(description="Exact quote from the memory supporting the observation")
+    relevance: str = Field(default="", description="Brief explanation of how this quote supports the observation")
+    timestamp: datetime = Field(description="When the source memory was created")
+
+    @field_validator("timestamp", mode="before")
+    @classmethod
+    def ensure_timezone_aware(cls, v: datetime | str | None) -> datetime:
+        """Ensure timestamp is always timezone-aware UTC."""
+        if v is None:
+            return datetime.now(timezone.utc)
+        if isinstance(v, str):
+            # Parse ISO format string, handling 'Z' suffix
+            v = datetime.fromisoformat(v.replace("Z", "+00:00"))
+        if isinstance(v, datetime):
+            if v.tzinfo is None:
+                return v.replace(tzinfo=timezone.utc)
+            return v
+        raise ValueError(f"Invalid timestamp type: {type(v)}")
+
+
+class Observation(BaseModel):
+    """A single observation within a mental model.
+
+    Observations represent patterns, preferences, beliefs, or other insights
+    derived from memories. Each observation must be grounded in evidence
+    with exact quotes from source memories.
+    """
+
+    title: str = Field(description="Short summary title for the observation (5-10 words)")
+    content: str = Field(description="The observation content - detailed explanation of what we believe to be true")
+    evidence: list[ObservationEvidence] = Field(default_factory=list, description="Supporting evidence with quotes")
+    created_at: datetime = Field(
+        default_factory=lambda: datetime.now(timezone.utc), description="When this observation was first created"
+    )
+
+    @field_validator("created_at", mode="before")
+    @classmethod
+    def ensure_created_at_timezone_aware(cls, v: datetime | str | None) -> datetime:
+        """Ensure created_at is always timezone-aware UTC."""
+        if v is None:
+            return datetime.now(timezone.utc)
+        if isinstance(v, str):
+            v = datetime.fromisoformat(v.replace("Z", "+00:00"))
+        if isinstance(v, datetime):
+            if v.tzinfo is None:
+                return v.replace(tzinfo=timezone.utc)
+            return v
+        raise ValueError(f"Invalid created_at type: {type(v)}")
+
+    @computed_field
+    @property
+    def trend(self) -> Trend:
+        """Compute trend from evidence timestamps."""
+        return compute_trend(self.evidence)
+
+    @computed_field
+    @property
+    def evidence_span(self) -> dict[str, str | None]:
+        """Get the time span covered by evidence."""
+        if not self.evidence:
+            return {"from": None, "to": None}
+        timestamps = [e.timestamp for e in self.evidence]
+        return {
+            "from": min(timestamps).isoformat(),
+            "to": max(timestamps).isoformat(),
+        }
+
+    @computed_field
+    @property
+    def evidence_count(self) -> int:
+        """Number of evidence items supporting this observation."""
+        return len(self.evidence)
+
+
+def compute_trend(
+    evidence: list[ObservationEvidence],
+    now: datetime | None = None,
+    recent_days: int = 30,
+    old_days: int = 90,
+) -> Trend:
+    """Compute the trend for an observation based on evidence timestamps.
+
+    The trend indicates how the evidence is distributed over time:
+    - STABLE: Evidence spread across time, continues to present
+    - STRENGTHENING: More evidence recently than historically
+    - WEAKENING: Evidence mostly old, sparse recently
+    - NEW: All evidence is recent (within recent_days)
+    - STALE: No evidence in recent window
+
+    Args:
+        evidence: List of evidence items with timestamps
+        now: Reference time for calculations (defaults to current UTC time)
+        recent_days: Number of days to consider "recent" (default 30)
+        old_days: Number of days to consider "old" (default 90)
+
+    Returns:
+        Computed Trend enum value
+    """
+    if now is None:
+        now = datetime.now(timezone.utc)
+
+    # Ensure now is timezone-aware
+    if now.tzinfo is None:
+        now = now.replace(tzinfo=timezone.utc)
+
+    if not evidence:
+        return Trend.STALE
+
+    recent_cutoff = now - timedelta(days=recent_days)
+    old_cutoff = now - timedelta(days=old_days)
+
+    # Normalize timestamps to UTC for comparison
+    def normalize_ts(ts: datetime) -> datetime:
+        if ts.tzinfo is None:
+            return ts.replace(tzinfo=timezone.utc)
+        return ts
+
+    recent = [e for e in evidence if normalize_ts(e.timestamp) > recent_cutoff]
+    old = [e for e in evidence if normalize_ts(e.timestamp) < old_cutoff]
+    middle = [e for e in evidence if old_cutoff <= normalize_ts(e.timestamp) <= recent_cutoff]
+
+    # No recent evidence = stale
+    if not recent:
+        return Trend.STALE
+
+    # All evidence is recent = new
+    if not old and not middle:
+        return Trend.NEW
+
+    # Compare density (evidence per day)
+    recent_density = len(recent) / recent_days if recent_days > 0 else 0
+    older_period = old_days - recent_days
+    older_density = (len(old) + len(middle)) / older_period if older_period > 0 else 0
+
+    # Avoid division by zero
+    if older_density == 0:
+        return Trend.NEW
+
+    ratio = recent_density / older_density
+
+    if ratio > 1.5:
+        return Trend.STRENGTHENING
+    elif ratio < 0.5:
+        return Trend.WEAKENING
+    else:
+        return Trend.STABLE