docs: add FAQ entry for conversation retain format

Addresses common questions from community discussions on the recommended format and flow for retaining conversations (JSON array vs plain text, upsert pattern, avoiding pre-summarization).
Fix Iris parser httpx read timeout for file uploads (#524 )
2026-03-09 15:05:19 +01:00 · 2026-03-09 11:22:44 +01:00 · 2026-03-09 11:22:24 +01:00 · 2026-03-07 09:00:13 +01:00 · 2026-03-07 08:59:51 +01:00 · 2026-03-06 12:35:12 -05:00
892 changed files with 167092 additions and 14917 deletions
@@ -2,10 +2,10 @@
 # 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
+HINDSIGHT_API_LLM_MODEL=gpt-4o-mini
 HINDSIGHT_API_LLM_BASE_URL=https://api.openai.com/v1

 # Example: Anthropic Claude configuration
@@ -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
@@ -31,6 +31,9 @@ jobs:
      - run: npm ci --workspace=hindsight-docs
      - run: uv run generate-llms-full
      - run: npm run build --workspace=hindsight-docs
+        env:
+          UMAMI_URL: https://analytics.hindsight.vectorize.io
+          UMAMI_WEBSITE_ID: ${{ secrets.UMAMI_WEBSITE_ID }}
      - uses: actions/upload-pages-artifact@v3
        with:
          path: hindsight-docs/build
@@ -46,6 +46,14 @@ jobs:
      working-directory: ./hindsight-embed
      run: uv build --out-dir dist

+    - name: Build hindsight-crewai
+      working-directory: ./hindsight-integrations/crewai
+      run: uv build --out-dir dist
+
+    - name: Build hindsight-pydantic-ai
+      working-directory: ./hindsight-integrations/pydantic-ai
+      run: uv build --out-dir dist
+
    # Publish in order (client and api first, then hindsight-all which depends on them)
    - name: Publish hindsight-client to PyPI
      uses: pypa/gh-action-pypi-publish@release/v1
@@ -77,6 +85,18 @@ jobs:
        packages-dir: ./hindsight-embed/dist
        skip-existing: true

+    - name: Publish hindsight-crewai to PyPI
+      uses: pypa/gh-action-pypi-publish@release/v1
+      with:
+        packages-dir: ./hindsight-integrations/crewai/dist
+        skip-existing: true
+
+    - name: Publish hindsight-pydantic-ai to PyPI
+      uses: pypa/gh-action-pypi-publish@release/v1
+      with:
+        packages-dir: ./hindsight-integrations/pydantic-ai/dist
+        skip-existing: true
+
    # Upload artifacts for GitHub release
    - name: Upload artifacts
      uses: actions/upload-artifact@v4
@@ -88,6 +108,8 @@ jobs:
          hindsight/dist/*
          hindsight-integrations/litellm/dist/*
          hindsight-embed/dist/*
+          hindsight-integrations/crewai/dist/*
+          hindsight-integrations/pydantic-ai/dist/*
        retention-days: 1

  release-typescript-client:
@@ -139,6 +161,153 @@ 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: ${{ secrets.NPM_TOKEN }}
+
+    - name: Pack for GitHub release
+      working-directory: ./hindsight-integrations/ai-sdk
+      run: npm pack
+
+    - name: Upload artifacts
+      uses: actions/upload-artifact@v4
+      with:
+        name: ai-sdk-integration
+        path: hindsight-integrations/ai-sdk/*.tgz
+        retention-days: 1
+
+  release-chat-integration:
+    runs-on: ubuntu-latest
+    environment: npm
+
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Set up Node.js
+      uses: actions/setup-node@v4
+      with:
+        node-version: '22'
+        registry-url: 'https://registry.npmjs.org'
+
+    - name: Install dependencies
+      working-directory: ./hindsight-integrations/chat
+      run: npm ci
+
+    - name: Build
+      working-directory: ./hindsight-integrations/chat
+      run: npm run build
+
+    - name: Publish to npm
+      working-directory: ./hindsight-integrations/chat
+      run: |
+        set +e
+        OUTPUT=$(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/chat
+      run: npm pack
+
+    - name: Upload artifacts
+      uses: actions/upload-artifact@v4
+      with:
+        name: chat-integration
+        path: hindsight-integrations/chat/*.tgz
+        retention-days: 1
+
  release-control-plane:
    runs-on: ubuntu-latest
    environment: npm
@@ -170,11 +339,14 @@ jobs:
    - name: Build
      run: npm run build --workspace=hindsight-control-plane

+    - name: Verify standalone build
+      run: test -f hindsight-control-plane/standalone/server.js || (echo 'standalone/server.js missing - build failed' && exit 1)
+
    - name: Publish to npm
      working-directory: ./hindsight-control-plane
      run: |
        set +e
-        OUTPUT=$(npm publish --access public 2>&1)
+        OUTPUT=$(npm publish --access public --ignore-scripts 2>&1)
        EXIT_CODE=$?
        echo "$OUTPUT"
        if [ $EXIT_CODE -ne 0 ]; then
@@ -215,6 +387,10 @@ jobs:
            target: aarch64-apple-darwin
            artifact_name: hindsight
            asset_name: hindsight-darwin-arm64
+          - os: ubuntu-24.04-arm
+            target: aarch64-unknown-linux-gnu
+            artifact_name: hindsight
+            asset_name: hindsight-linux-arm64

    steps:
    - uses: actions/checkout@v4
@@ -242,6 +418,7 @@ 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 +428,28 @@ 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 +487,9 @@ 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 +515,7 @@ 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 +524,7 @@ 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 +565,7 @@ 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-chat-integration, release-control-plane, release-rust-cli, release-docker-images, release-helm-chart]
    permissions:
      contents: write

@@ -389,6 +588,24 @@ 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 Chat Integration
+      uses: actions/download-artifact@v4
+      with:
+        name: chat-integration
+        path: ./artifacts/chat-integration
+
    - name: Download Control Plane
      uses: actions/download-artifact@v4
      with:
@@ -427,9 +644,16 @@ jobs:
        cp artifacts/python-packages/hindsight-api/dist/* release-assets/ || true
        cp artifacts/python-packages/hindsight/dist/* release-assets/ || true
        cp artifacts/python-packages/hindsight-integrations/litellm/dist/* release-assets/ || true
+        cp artifacts/python-packages/hindsight-integrations/pydantic-ai/dist/* release-assets/ || true
        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
+        # Chat Integration
+        cp artifacts/chat-integration/*.tgz release-assets/ || true
        # Control Plane
        cp artifacts/control-plane/*.tgz release-assets/ || true
        # Rust CLI binaries
@@ -45,9 +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*
@@ -45,6 +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)
@@ -56,8 +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
 ```

@@ -237,26 +245,61 @@ def process(data: UserData) -> str:

 ### 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

@@ -274,9 +317,10 @@ npm install
 Required env vars:
 - `HINDSIGHT_API_LLM_PROVIDER`: openai, anthropic, gemini, groq, ollama, lmstudio
 - `HINDSIGHT_API_LLM_API_KEY`: Your API key
- `HINDSIGHT_API_LLM_MODEL`: Model name (e.g., o3-mini, claude-sonnet-4-20250514)
+- `HINDSIGHT_API_LLM_MODEL`: Model name (e.g., gpt-4o-mini, claude-sonnet-4-20250514)

 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: true)
@@ -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:
@@ -1,8 +1,8 @@
 <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-3nhbm4w29-LeSJ5Ixi6j8PdiYOCPlOgg)
@@ -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 to 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 implementations 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:

@@ -262,3 +307,5 @@ MIT — see [LICENSE](./LICENSE)
 ---

 Built by [Vectorize.io](https://vectorize.io)
+
+<img src="https://umami-pixel.chris-latimer.workers.dev/?id=a8b043e6-6964-454d-80df-69b69d3f0d50&host=github.com&url=/vectorize-io/hindsight" width="1" height="1" alt="" />
@@ -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:
@@ -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.
@@ -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:
@@ -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;
+        }
+    }
+}
@@ -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
@@ -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:
@@ -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:
@@ -0,0 +1,19 @@
+{
+  "identities": [
+    {
+      "name": "hindsight",
+      "credentials": [
+        {
+          "accessKey": "hindsight_s3_key",
+          "secretKey": "hindsight_s3_secret"
+        }
+      ],
+      "actions": [
+        "Admin",
+        "Read",
+        "Write",
+        "List"
+      ]
+    }
+  ]
+}
@@ -0,0 +1,16 @@
+# Git
+.git
+.gitignore
+.gitattributes
+
+# Docker
+docker-compose.yaml
+.dockerignore
+
+# Documentation
+README.md
+*.md
+
+# Environment
+.env
+.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
@@ -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
+
@@ -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)
@@ -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:
@@ -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:
@@ -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
@@ -165,20 +170,64 @@ RUN chown -R hindsight:hindsight /app

 USER hindsight

+# Create pg0 data directory as hindsight user so that Docker seeds new named
+# volumes with correct ownership (UID 1000) on first use, avoiding the
+# "Permission denied" error when mounting a fresh root-owned volume.
+RUN mkdir -p /home/hindsight/.pg0
+
 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 +239,10 @@ 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"]

@@ -273,20 +326,64 @@ RUN chown -R hindsight:hindsight /app

 USER hindsight

+# Create pg0 data directory as hindsight user so that Docker seeds new named
+# volumes with correct ownership (UID 1000) on first use, avoiding the
+# "Permission denied" error when mounting a fresh root-owned volume.
+RUN mkdir -p /home/hindsight/.pg0
+
 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 +397,10 @@ 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"]

@@ -97,7 +97,7 @@ fi
 if [ "$ENABLE_CP" = "true" ]; then
    echo "🎛️  Starting Control Plane..."
    cd /app/control-plane
-    PORT=9999 node server.js &
+    PORT="${HINDSIGHT_CP_PORT:-9999}" node server.js &
    CP_PID=$!
    PIDS+=($CP_PID)
 else
@@ -110,7 +110,7 @@ echo "✅ Hindsight is running!"
 echo ""
 echo "📍 Access:"
 if [ "$ENABLE_CP" = "true" ]; then
-    echo "   Control Plane: http://localhost:9999"
+    echo "   Control Plane: http://localhost:${HINDSIGHT_CP_PORT:-9999}"
 fi
 if [ "$ENABLE_API" = "true" ]; then
    echo "   API:           http://localhost:8888"
@@ -0,0 +1,217 @@
+#!/bin/bash
+#
+# Docker Smoke Test Script
+#
+# Tests that a Hindsight Docker image starts correctly and becomes healthy.
+# Can be run locally or in CI pipelines.
+#
+# Usage:
+#   ./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:
+#   HINDSIGHT_API_LLM_API_KEY                   - Required for API/standalone images (LLM verification)
+#   HINDSIGHT_API_LLM_PROVIDER                  - LLM provider (default: openai)
+#   HINDSIGHT_API_LLM_MODEL                     - LLM model (default: gpt-4o-mini)
+#   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 full image
+#   ./docker/test-image.sh hindsight-api:test
+#
+#   # Test a released image
+#   ./docker/test-image.sh ghcr.io/vectorize-io/hindsight:latest
+#
+#   # Test control plane image
+#   ./docker/test-image.sh hindsight-control-plane:test cp-only
+#
+#   # Test slim image with external providers
+#   export HINDSIGHT_API_LLM_API_KEY=sk_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)
+#   1 - Failure (container not healthy within timeout)
+#   2 - Invalid arguments
+#
+
+set -euo pipefail
+
+# Colors for output
+RED='\033[0;31m'
+GREEN='\033[0;32m'
+YELLOW='\033[0;33m'
+NC='\033[0m' # No Color
+
+# Configuration
+IMAGE="${1:-}"
+TARGET="${2:-api}"
+TIMEOUT="${SMOKE_TEST_TIMEOUT:-120}"
+CONTAINER_NAME="${SMOKE_TEST_CONTAINER_NAME:-hindsight-smoke-test}"
+LLM_PROVIDER="${HINDSIGHT_API_LLM_PROVIDER:-openai}"
+LLM_MODEL="${HINDSIGHT_API_LLM_MODEL:-gpt-4o-mini}"
+
+# Validate arguments
+if [ -z "$IMAGE" ]; then
+    echo -e "${RED}Error: Image argument is required${NC}"
+    echo ""
+    echo "Usage: $0 <image> [target]"
+    echo ""
+    echo "Examples:"
+    echo "  $0 hindsight-api:test"
+    echo "  $0 ghcr.io/vectorize-io/hindsight:latest"
+    echo "  $0 hindsight-control-plane:test cp-only"
+    exit 2
+fi
+
+# Determine health endpoint based on target
+if [ "$TARGET" = "cp-only" ]; then
+    HEALTH_PORT=9999
+    HEALTH_PATH="/api/health"
+    NEEDS_LLM=false
+else
+    HEALTH_PORT=8888
+    HEALTH_PATH="/health"
+    NEEDS_LLM=true
+fi
+
+# Check for required environment variables
+if [ "$NEEDS_LLM" = true ] && [ "$LLM_PROVIDER" != "vertexai" ] && [ -z "${HINDSIGHT_API_LLM_API_KEY:-}" ]; then
+    echo -e "${RED}Error: HINDSIGHT_API_LLM_API_KEY environment variable is required for API/standalone images${NC}"
+    echo "Set it with: export HINDSIGHT_API_LLM_API_KEY=your-api-key"
+    exit 2
+fi
+
+# Cleanup function
+cleanup() {
+    echo "Cleaning up..."
+    docker stop "$CONTAINER_NAME" 2>/dev/null || true
+    docker rm "$CONTAINER_NAME" 2>/dev/null || true
+}
+
+# Set trap to cleanup on exit
+trap cleanup EXIT
+
+echo -e "${YELLOW}Starting smoke test for: ${IMAGE}${NC}"
+echo "  Target: $TARGET"
+echo "  Health endpoint: http://localhost:${HEALTH_PORT}${HEALTH_PATH}"
+echo "  Timeout: ${TIMEOUT}s"
+echo ""
+
+# Remove any existing container with the same name
+docker rm -f "$CONTAINER_NAME" 2>/dev/null || true
+
+# Start container based on target type
+echo "Starting container..."
+if [ "$TARGET" = "cp-only" ]; then
+    docker run -d --name "$CONTAINER_NAME" \
+        -p "${HEALTH_PORT}:${HEALTH_PORT}" \
+        "$IMAGE"
+else
+    # 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=$LLM_PROVIDER"
+    if [ -n "${HINDSIGHT_API_LLM_API_KEY:-}" ]; then
+        DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_LLM_API_KEY=${HINDSIGHT_API_LLM_API_KEY}"
+    fi
+    DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_LLM_MODEL=$LLM_MODEL"
+
+    # Add Vertex AI config if provider is vertexai
+    if [ "$LLM_PROVIDER" = "vertexai" ]; then
+        if [ -n "${HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY:-}" ]; then
+            DOCKER_CMD="$DOCKER_CMD -v ${HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY}:/tmp/gcp-credentials.json:ro"
+            DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY=/tmp/gcp-credentials.json"
+        fi
+        if [ -n "${HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID:-}" ]; then
+            DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID=${HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID}"
+        fi
+        if [ -n "${HINDSIGHT_API_LLM_VERTEXAI_REGION:-}" ]; then
+            DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_LLM_VERTEXAI_REGION=${HINDSIGHT_API_LLM_VERTEXAI_REGION}"
+        fi
+    fi
+
+    # 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 $DOCKER_CMD
+fi
+
+# Wait for health endpoint
+echo "Waiting for health endpoint at http://localhost:${HEALTH_PORT}${HEALTH_PATH}..."
+start_time=$(date +%s)
+
+for i in $(seq 1 "$TIMEOUT"); do
+    if curl -sf "http://localhost:${HEALTH_PORT}${HEALTH_PATH}" > /dev/null 2>&1; then
+        end_time=$(date +%s)
+        duration=$((end_time - start_time))
+        echo ""
+        echo -e "${GREEN}Container is healthy after ${duration}s${NC}"
+        echo ""
+        echo "=== Health Response ==="
+        curl -s "http://localhost:${HEALTH_PORT}${HEALTH_PATH}" | python3 -m json.tool 2>/dev/null || curl -s "http://localhost:${HEALTH_PORT}${HEALTH_PATH}"
+        echo ""
+        echo ""
+        echo "=== Container Logs (last 50 lines) ==="
+        docker logs "$CONTAINER_NAME" 2>&1 | tail -50
+        echo ""
+        echo -e "${GREEN}Smoke test PASSED${NC}"
+        exit 0
+    fi
+
+    # Show progress every 10 seconds
+    if [ $((i % 10)) -eq 0 ]; then
+        echo "  Still waiting... (${i}s)"
+    fi
+
+    # Check if container is still running
+    if ! docker ps -q -f "name=$CONTAINER_NAME" | grep -q .; then
+        echo ""
+        echo -e "${RED}Container exited unexpectedly!${NC}"
+        echo ""
+        echo "=== Container Logs ==="
+        docker logs "$CONTAINER_NAME" 2>&1
+        echo ""
+        echo -e "${RED}Smoke test FAILED${NC}"
+        exit 1
+    fi
+
+    sleep 1
+done
+
+# Timeout reached
+echo ""
+echo -e "${RED}Container failed to become healthy after ${TIMEOUT}s${NC}"
+echo ""
+echo "=== Container Logs ==="
+docker logs "$CONTAINER_NAME" 2>&1
+echo ""
+echo -e "${RED}Smoke test FAILED${NC}"
+exit 1
@@ -0,0 +1,47 @@
+#!/bin/bash
+#
+# Local Test Script for Slim Docker Images
+#
+# This script makes it easy to test slim images locally with external providers.
+# It expects API keys to be set in environment variables.
+#
+# Usage:
+#   export OPENAI_API_KEY=sk-xxx
+#   export COHERE_API_KEY=xxx
+#   ./docker/test-slim-local.sh
+#
+# Or inline:
+#   OPENAI_API_KEY=sk_xxx COHERE_API_KEY=xxx ./docker/test-slim-local.sh
+#
+
+set -euo pipefail
+
+# Check for required API keys
+if [ -z "${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: $IMAGE"
+echo ""
+
+# Set up LLM and external providers
+export HINDSIGHT_API_LLM_PROVIDER=openai
+export HINDSIGHT_API_LLM_API_KEY=$OPENAI_API_KEY
+export HINDSIGHT_API_LLM_MODEL=gpt-4o-mini
+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
@@ -2,8 +2,8 @@ apiVersion: v2
 name: hindsight
 description: Hindsight helm chart
 type: application
-version: 0.3.0
-appVersion: "0.3.0"
+version: 0.4.16
+appVersion: "0.4.16"
 keywords:
  - ai
  - memory
@@ -127,6 +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
 */}}
@@ -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
@@ -60,10 +60,25 @@ spec:
        - 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 }}
@@ -84,7 +99,7 @@ spec:
      nodeSelector:
        {{- toYaml . | nindent 8 }}
      {{- end }}
-      {{- with .Values.affinity }}
+      {{- with (.Values.api.affinity | default .Values.affinity) }}
      affinity:
        {{- toYaml . | nindent 8 }}
      {{- end }}
@@ -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 }}
@@ -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 }}
@@ -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 }}
@@ -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 }}
@@ -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 }}
@@ -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 }}
@@ -32,7 +32,7 @@ spec:
      - name: worker
        securityContext:
          {{- toYaml .Values.securityContext | nindent 10 }}
-        image: "{{ .Values.worker.image.repository }}:{{ .Values.worker.image.tag | default .Values.version }}"
+        image: "{{ .Values.worker.image.repository }}:{{ .Values.worker.image.tag | default .Values.version | default .Chart.AppVersion }}"
        imagePullPolicy: {{ .Values.worker.image.pullPolicy }}
        command: ["hindsight-worker"]
        ports:
@@ -99,7 +99,7 @@ spec:
      nodeSelector:
        {{- toYaml . | nindent 8 }}
      {{- end }}
-      {{- with .Values.affinity }}
+      {{- with (.Values.worker.affinity | default .Values.affinity) }}
      affinity:
        {{- toYaml . | nindent 8 }}
      {{- end }}
@@ -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"
@@ -75,7 +85,7 @@ worker:
  image:
    repository: ghcr.io/vectorize-io/hindsight-api
    pullPolicy: IfNotPresent
-    # tag defaults to .Values.version if not specified
+    # tag: ""  # defaults to .Values.version, then Chart.appVersion if not specified

  service:
    # Service for metrics scraping (headless for StatefulSet)
@@ -121,6 +131,15 @@ worker:
    # 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: {}

@@ -164,6 +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"
@@ -262,9 +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
@@ -46,4 +46,4 @@ __all__ = [
    "RemoteTEICrossEncoder",
    "LLMConfig",
 ]
-__version__ = "0.1.0"
+__version__ = "0.4.16"
@@ -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
@@ -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")
@@ -0,0 +1,88 @@
+"""Add text_signals column to memory_units for enriched BM25 indexing.
+
+text_signals stores a denormalized space-separated string of entity names
+(and future signals) to improve full-text search recall without polluting
+the stored fact text.
+
+- vchord: text_signals included in tokenize() at insert time
+- native: search_vector GENERATED column regenerated to include text_signals
+- pg_textsearch: no change (index only supports a single base column)
+
+Revision ID: a2b3c4d5e6f7
+Revises: z1u2v3w4x5y6
+Create Date: 2026-02-28
+"""
+
+import os
+from collections.abc import Sequence
+
+from alembic import context, op
+
+revision: str = "a2b3c4d5e6f7"
+down_revision: str | Sequence[str] | None = "aa2b3c4d5e6f"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def _detect_text_search_extension() -> str:
+    return os.getenv("HINDSIGHT_API_TEXT_SEARCH_EXTENSION", "native").lower()
+
+
+def upgrade() -> None:
+    schema = _get_schema_prefix()
+    table = f"{schema}memory_units"
+    text_search_ext = _detect_text_search_extension()
+
+    # Add text_signals column (nullable TEXT, populated at retain time)
+    op.execute(f"ALTER TABLE {table} ADD COLUMN IF NOT EXISTS text_signals TEXT")
+
+    if text_search_ext == "native":
+        # Native PostgreSQL: drop and recreate the GENERATED tsvector column to include text_signals
+        op.execute(f"ALTER TABLE {table} DROP COLUMN IF EXISTS search_vector")
+        op.execute(f"""
+            ALTER TABLE {table}
+            ADD COLUMN search_vector tsvector
+            GENERATED ALWAYS AS (
+                to_tsvector('english',
+                    COALESCE(text, '') || ' ' ||
+                    COALESCE(context, '') || ' ' ||
+                    COALESCE(text_signals, '')
+                )
+            ) STORED
+        """)
+        # Recreate GIN index (was dropped with the column)
+        op.execute(f"""
+            CREATE INDEX IF NOT EXISTS idx_memory_units_text_search
+            ON {table} USING gin(search_vector)
+        """)
+
+    # vchord: tokenize() call in fact_storage.py is updated to include text_signals at insert time
+    # pg_textsearch: no change — index operates on the base `text` column only
+
+
+def downgrade() -> None:
+    schema = _get_schema_prefix()
+    table = f"{schema}memory_units"
+    text_search_ext = _detect_text_search_extension()
+
+    if text_search_ext == "native":
+        op.execute(f"DROP INDEX IF EXISTS {schema}idx_memory_units_text_search")
+        op.execute(f"ALTER TABLE {table} DROP COLUMN IF EXISTS search_vector")
+        op.execute(f"""
+            ALTER TABLE {table}
+            ADD COLUMN search_vector tsvector
+            GENERATED ALWAYS AS (
+                to_tsvector('english', COALESCE(text, '') || ' ' || COALESCE(context, ''))
+            ) STORED
+        """)
+        op.execute(f"""
+            CREATE INDEX idx_memory_units_text_search
+            ON {table} USING gin(search_vector)
+        """)
+
+    op.execute(f"ALTER TABLE {table} DROP COLUMN IF EXISTS text_signals")
@@ -0,0 +1,54 @@
+"""Add GIN index on source_memory_ids for observation lookup performance
+
+Without this index, queries using the array overlap operator (&&) or array
+containment (@>) on source_memory_ids require a full sequential scan over all
+observation memory_units. At ~77k observations this was measured at 45ms per
+query, becoming a bottleneck during consolidation recall (57-64s timeouts) and
+user recall (18-27s average).
+
+The GIN index reduces these queries to index scans: 45ms → 0.049ms (927x
+speedup). Recall dropped from 18-27s to ~6s, and consolidation recall
+stabilised from timeout to ~15s.
+
+Created with CONCURRENTLY so the migration does not block reads or writes.
+CONCURRENTLY requires running outside a transaction block, so the migration
+emits an explicit COMMIT before the statement and uses IF NOT EXISTS for
+idempotency.
+
+Revision ID: a2b3c4d5e6f8
+Revises: f7g8h9i0j1k2
+Create Date: 2026-03-04
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+revision: str = "a2b3c4d5e6f8"
+down_revision: str | Sequence[str] | None = "f7g8h9i0j1k2"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    schema = _get_schema_prefix()
+
+    # CREATE INDEX CONCURRENTLY cannot run inside a transaction block.
+    # Commit the current Alembic transaction first.
+    op.execute("COMMIT")
+    op.execute(
+        f"CREATE INDEX CONCURRENTLY IF NOT EXISTS idx_memory_units_source_memory_ids "
+        f"ON {schema}memory_units USING GIN (source_memory_ids) "
+        f"WHERE source_memory_ids IS NOT NULL"
+    )
+
+
+def downgrade() -> None:
+    schema = _get_schema_prefix()
+    op.execute("COMMIT")
+    op.execute(f"DROP INDEX CONCURRENTLY IF EXISTS {schema}idx_memory_units_source_memory_ids")
@@ -0,0 +1,36 @@
+"""Make event_date nullable in memory_units to support timestamp-free content
+
+Revision ID: aa2b3c4d5e6f
+Revises: z1u2v3w4x5y6
+Create Date: 2026-03-02
+
+When callers retain content without a timestamp (e.g. fictional documents, static text),
+the event_date column should be allowed to be NULL rather than defaulting to utcnow().
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+revision: str = "aa2b3c4d5e6f"
+down_revision: str | Sequence[str] | None = "z1u2v3w4x5y6"
+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"ALTER TABLE {schema}memory_units ALTER COLUMN event_date DROP NOT NULL")
+
+
+def downgrade() -> None:
+    schema = _get_schema_prefix()
+    # Backfill NULLs with now() before restoring the NOT NULL constraint
+    op.execute(f"UPDATE {schema}memory_units SET event_date = now() WHERE event_date IS NULL")
+    op.execute(f"ALTER TABLE {schema}memory_units ALTER COLUMN event_date SET NOT NULL")
@@ -0,0 +1,68 @@
+"""Add partial indexes on memory_units temporal date fields for fast temporal retrieval
+
+Revision ID: b3c4d5e6f7g8
+Revises: c1a2b3d4e5f6
+Create Date: 2026-03-02
+
+The temporal retrieval entry-point query filters memory_units by occurred_start,
+occurred_end, and mentioned_at using OR conditions. Without dedicated indexes the
+planner falls back to a sequential scan of all bank rows after applying the
+(bank_id, fact_type) index, then re-checks each date field.
+
+These three partial indexes give the planner bitmap-index scan options for the
+three most common date predicates, dramatically reducing the row set before any
+embedding computation is required.
+
+All indexes are created CONCURRENTLY so the migration does not block writes on
+memory_units during production deployments. CONCURRENTLY requires running outside
+a transaction block; see migrations.py for how this is handled safely.
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+revision: str = "b3c4d5e6f7g8"
+down_revision: str | Sequence[str] | None = "c1a2b3d4e5f6"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    schema = _get_schema_prefix()
+    # Partial index on occurred_start (covers "occurred_start BETWEEN $4 AND $5")
+    op.execute("COMMIT")
+    op.execute(
+        f"CREATE INDEX CONCURRENTLY IF NOT EXISTS idx_memory_units_bank_occurred_start "
+        f"ON {schema}memory_units(bank_id, fact_type, occurred_start) "
+        f"WHERE occurred_start IS NOT NULL"
+    )
+    # Partial index on occurred_end (covers "occurred_end BETWEEN $4 AND $5")
+    op.execute("COMMIT")
+    op.execute(
+        f"CREATE INDEX CONCURRENTLY IF NOT EXISTS idx_memory_units_bank_occurred_end "
+        f"ON {schema}memory_units(bank_id, fact_type, occurred_end) "
+        f"WHERE occurred_end IS NOT NULL"
+    )
+    # Partial index on mentioned_at (covers "mentioned_at BETWEEN $4 AND $5")
+    op.execute("COMMIT")
+    op.execute(
+        f"CREATE INDEX CONCURRENTLY IF NOT EXISTS idx_memory_units_bank_mentioned_at "
+        f"ON {schema}memory_units(bank_id, fact_type, mentioned_at) "
+        f"WHERE mentioned_at IS NOT NULL"
+    )
+
+
+def downgrade() -> None:
+    schema = _get_schema_prefix()
+    op.execute("COMMIT")
+    op.execute(f"DROP INDEX CONCURRENTLY IF EXISTS {schema}idx_memory_units_bank_mentioned_at")
+    op.execute("COMMIT")
+    op.execute(f"DROP INDEX CONCURRENTLY IF EXISTS {schema}idx_memory_units_bank_occurred_end")
+    op.execute("COMMIT")
+    op.execute(f"DROP INDEX CONCURRENTLY IF EXISTS {schema}idx_memory_units_bank_occurred_start")
@@ -0,0 +1,34 @@
+"""Backfill observation_scopes column if missing.
+
+This migration ensures observation_scopes exists even on databases that had
+revision z1u2v3w4x5y6 applied when it referred to the old text_signals migration
+(before it was renamed to a2b3c4d5e6f7). The ADD COLUMN IF NOT EXISTS makes this
+a no-op on databases that already have the column.
+
+Revision ID: b4c5d6e7f8a9
+Revises: a2b3c4d5e6f7
+Create Date: 2026-03-02
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+revision: str = "b4c5d6e7f8a9"
+down_revision: str | Sequence[str] | None = "a2b3c4d5e6f7"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    schema = _get_schema_prefix()
+    op.execute(f"ALTER TABLE {schema}memory_units ADD COLUMN IF NOT EXISTS observation_scopes JSONB")
+
+
+def downgrade() -> None:
+    pass  # intentionally no-op — safe to leave the column in place
@@ -0,0 +1,46 @@
+"""Enable pg_trgm extension and add GIN trigram index on entities.canonical_name
+
+Revision ID: c1a2b3d4e5f6
+Revises: b4c5d6e7f8a9
+Create Date: 2026-03-02
+
+Index is created CONCURRENTLY so the migration does not block writes on entities
+during production deployments. CONCURRENTLY requires running outside a transaction
+block; see migrations.py for how this is handled safely.
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+revision: str = "c1a2b3d4e5f6"
+down_revision: str | Sequence[str] | None = "b4c5d6e7f8a9"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    # pg_trgm ships with every standard PostgreSQL installation as a contrib module.
+    # It enables fast similarity lookups via GIN indexes, used for entity name matching.
+    op.execute("CREATE EXTENSION IF NOT EXISTS pg_trgm")
+
+    schema = _get_schema_prefix()
+    # GIN index on canonical_name enables sub-millisecond trigram similarity queries
+    # (% operator, similarity()) instead of full-table scans across all bank entities.
+    op.execute("COMMIT")
+    op.execute(
+        f"CREATE INDEX CONCURRENTLY IF NOT EXISTS entities_canonical_name_trgm_idx "
+        f"ON {schema}entities USING GIN (canonical_name gin_trgm_ops)"
+    )
+
+
+def downgrade() -> None:
+    schema = _get_schema_prefix()
+    op.execute("COMMIT")
+    op.execute(f"DROP INDEX CONCURRENTLY IF EXISTS {schema}entities_canonical_name_trgm_idx")
+    # Note: not dropping pg_trgm extension as other indexes may depend on it
@@ -0,0 +1,30 @@
+"""Add history column to mental_models
+
+Revision ID: c3d4e5f6g7h8
+Revises: a2b3c4d5e6f7, a2b3c4d5e6f8
+Create Date: 2026-03-06
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+revision: str = "c3d4e5f6g7h8"
+down_revision: str | Sequence[str] | None = ("a2b3c4d5e6f7", "a2b3c4d5e6f8")
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    schema = _get_schema_prefix()
+    op.execute(f"ALTER TABLE {schema}mental_models ADD COLUMN IF NOT EXISTS history JSONB DEFAULT '[]'::jsonb")
+
+
+def downgrade() -> None:
+    schema = _get_schema_prefix()
+    op.execute(f"ALTER TABLE {schema}mental_models DROP COLUMN IF EXISTS history")
@@ -0,0 +1,83 @@
+"""Add covering and composite indexes to speed up link expansion graph retrieval.
+
+Two indexes target the two bottlenecks identified by EXPLAIN ANALYZE on a 17M-row
+memory_links table:
+
+1. idx_memory_links_to_type_weight  (to_unit_id, link_type, weight DESC)
+   The semantic incoming direction — finding facts that consider seeds as their
+   nearest neighbour — currently hits an expensive BitmapAnd of two separate
+   bitmap scans (to_unit_id bitmap ∩ link_type bitmap).  A composite index
+   on (to_unit_id, link_type) turns this into a single index scan and reduces
+   latency from ~36 ms to < 5 ms per query.
+
+2. idx_memory_links_entity_covering  (from_unit_id) INCLUDE (to_unit_id, entity_id)
+   WHERE link_type = 'entity'
+   The entity co-occurrence expansion uses COUNT(DISTINCT ml.entity_id) and
+   joins on ml.to_unit_id.  Without a covering index the planner must read
+   ~2 500 heap pages to fetch entity_id and to_unit_id after the bitmap index
+   scan, adding ~230 ms of random I/O.  INCLUDE adds those two columns to the
+   index leaf pages so the entire query can be served from the index (index-only
+   scan), eliminating the heap reads entirely.
+   Partial index (WHERE link_type = 'entity') keeps index size ~40 % smaller.
+
+Both indexes are created with CONCURRENTLY so the migration does not block
+concurrent reads or writes on memory_links.  CONCURRENTLY requires running
+outside a transaction block, so the migration emits an explicit COMMIT before
+each statement and uses IF NOT EXISTS for idempotency.
+
+Revision ID: d2e3f4a5b6c7
+Revises: b3c4d5e6f7g8
+Create Date: 2026-03-02
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+revision: str = "d2e3f4a5b6c7"
+down_revision: str | Sequence[str] | None = "b3c4d5e6f7g8"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    schema = _get_schema_prefix()
+
+    # CREATE INDEX CONCURRENTLY cannot run inside a transaction block.
+    # Commit the current Alembic transaction, then issue each CONCURRENTLY
+    # statement in its own implicit autocommit transaction.
+    # IF NOT EXISTS makes each statement idempotent if the migration is retried.
+
+    # Index for the semantic *incoming* direction in link_expansion_retrieval.py.
+    # Replaces the BitmapAnd of idx_memory_links_to_unit ∩ idx_memory_links_link_type
+    # with a single composite index scan.
+    op.execute("COMMIT")
+    op.execute(
+        f"CREATE INDEX CONCURRENTLY IF NOT EXISTS idx_memory_links_to_type_weight "
+        f"ON {schema}memory_links(to_unit_id, link_type, weight DESC)"
+    )
+
+    # Covering index for entity co-occurrence expansion.
+    # Enables an index-only scan: entity_id and to_unit_id are read from the
+    # index leaf pages instead of the heap, eliminating ~2 500 random heap-page
+    # reads per expansion query.
+    op.execute("COMMIT")
+    op.execute(
+        f"CREATE INDEX CONCURRENTLY IF NOT EXISTS idx_memory_links_entity_covering "
+        f"ON {schema}memory_links(from_unit_id) "
+        f"INCLUDE (to_unit_id, entity_id) "
+        f"WHERE link_type = 'entity'"
+    )
+
+
+def downgrade() -> None:
+    schema = _get_schema_prefix()
+    op.execute("COMMIT")
+    op.execute(f"DROP INDEX CONCURRENTLY IF EXISTS {schema}idx_memory_links_entity_covering")
+    op.execute("COMMIT")
+    op.execute(f"DROP INDEX CONCURRENTLY IF EXISTS {schema}idx_memory_links_to_type_weight")
@@ -0,0 +1,62 @@
+"""Add webhooks table and next_retry_at to async_operations.
+
+Webhook deliveries are handled as async_operations tasks (operation_type='webhook_delivery')
+rather than a dedicated webhook_deliveries table.
+
+Revision ID: e4f5a6b7c8d9
+Revises: d2e3f4a5b6c7
+Create Date: 2026-03-04
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+revision: str = "e4f5a6b7c8d9"
+down_revision: str | Sequence[str] | None = "d2e3f4a5b6c7"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    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 TABLE IF NOT EXISTS {schema}webhooks (
+            id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+            bank_id TEXT,
+            url TEXT NOT NULL,
+            secret TEXT,
+            event_types TEXT[] NOT NULL DEFAULT '{{}}',
+            enabled BOOLEAN NOT NULL DEFAULT TRUE,
+            created_at TIMESTAMPTZ NOT NULL DEFAULT NOW(),
+            updated_at TIMESTAMPTZ NOT NULL DEFAULT NOW()
+        )
+        """
+    )
+
+    # Index for bank-scoped webhook lookup
+    op.execute(f"CREATE INDEX IF NOT EXISTS idx_webhooks_bank_id ON {schema}webhooks(bank_id)")
+
+    # Add next_retry_at to async_operations for task-owned retry scheduling
+    op.execute(f"ALTER TABLE {schema}async_operations ADD COLUMN IF NOT EXISTS next_retry_at TIMESTAMPTZ NULL")
+
+    # Index for polling: status + next_retry_at
+    op.execute(
+        f"CREATE INDEX IF NOT EXISTS idx_async_operations_status_retry "
+        f"ON {schema}async_operations(status, next_retry_at)"
+    )
+
+
+def downgrade() -> None:
+    schema = _get_schema_prefix()
+    op.execute(f"DROP INDEX IF EXISTS {schema}idx_async_operations_status_retry")
+    op.execute(f"ALTER TABLE {schema}async_operations DROP COLUMN IF EXISTS next_retry_at")
+    op.execute(f"DROP INDEX IF EXISTS {schema}idx_webhooks_bank_id")
+    op.execute(f"DROP TABLE IF EXISTS {schema}webhooks")
@@ -0,0 +1,33 @@
+"""Add http_config JSONB column to webhooks table.
+
+Stores HTTP delivery configuration (method, timeout, headers, params) as a
+single JSONB column rather than separate columns.
+
+Revision ID: f7g8h9i0j1k2
+Revises: e4f5a6b7c8d9
+Create Date: 2026-03-04
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+revision: str = "f7g8h9i0j1k2"
+down_revision: str | Sequence[str] | None = "e4f5a6b7c8d9"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    schema = _get_schema_prefix()
+    op.execute(f"ALTER TABLE {schema}webhooks ADD COLUMN IF NOT EXISTS http_config JSONB NOT NULL DEFAULT '{{}}'")
+
+
+def downgrade() -> None:
+    schema = _get_schema_prefix()
+    op.execute(f"ALTER TABLE {schema}webhooks DROP COLUMN IF EXISTS http_config")
@@ -10,9 +10,11 @@ This migration:
 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"
@@ -27,10 +29,106 @@ def _get_schema_prefix() -> str:
    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 (
@@ -57,18 +155,60 @@ def upgrade() -> None:

    # Indexes for learnings
    op.execute(f"CREATE INDEX idx_learnings_bank_id ON {schema}learnings(bank_id)")
-    op.execute(f"""
-        CREATE INDEX idx_learnings_embedding ON {schema}learnings
-        USING hnsw (embedding vector_cosine_ops)
-    """)
+
+    # 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
-    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)")
+    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"""
@@ -94,21 +234,64 @@ def upgrade() -> None:

    # Indexes for pinned_reflections
    op.execute(f"CREATE INDEX idx_pinned_reflections_bank_id ON {schema}pinned_reflections(bank_id)")
-    op.execute(f"""
-        CREATE INDEX idx_pinned_reflections_embedding ON {schema}pinned_reflections
-        USING hnsw (embedding vector_cosine_ops)
-    """)
+
+    # 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
-    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)
-    """)
+    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"""
@@ -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)
+    """)
@@ -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
+    """)
@@ -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")
@@ -0,0 +1,35 @@
+"""Add observation_scopes column to memory_units table
+
+Revision ID: z1u2v3w4x5y6
+Revises: a1b2c3d4e5f6
+Create Date: 2026-02-25
+
+Adds observation_scopes JSONB column to memory_units to control how observations
+are scoped during consolidation. Accepts "per_tag", "combined", or an explicit
+list of tag-set lists for custom multi-pass consolidation.
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+revision: str = "z1u2v3w4x5y6"
+down_revision: str | Sequence[str] | None = "a1b2c3d4e5f6"
+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"ALTER TABLE {schema}memory_units ADD COLUMN IF NOT EXISTS observation_scopes JSONB")
+
+
+def downgrade() -> None:
+    schema = _get_schema_prefix()
+    op.execute(f"ALTER TABLE {schema}memory_units DROP COLUMN IF EXISTS observation_scopes")
@@ -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
@@ -8,7 +8,47 @@ from contextvars import ContextVar
 from fastmcp import FastMCP

 from hindsight_api import MemoryEngine
+from hindsight_api.config import _get_raw_config
+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
+
+# All tools available in the system (explicit list — no wildcards)
+_ALL_TOOLS: frozenset[str] = frozenset(
+    {
+        "retain",
+        "recall",
+        "reflect",
+        "list_banks",
+        "create_bank",
+        "list_mental_models",
+        "get_mental_model",
+        "create_mental_model",
+        "update_mental_model",
+        "delete_mental_model",
+        "refresh_mental_model",
+        "list_directives",
+        "create_directive",
+        "delete_directive",
+        "list_memories",
+        "get_memory",
+        "delete_memory",
+        "list_documents",
+        "get_document",
+        "delete_document",
+        "list_operations",
+        "get_operation",
+        "cancel_operation",
+        "list_tags",
+        "get_bank",
+        "get_bank_stats",
+        "update_bank",
+        "delete_bank",
+        "clear_memories",
+    }
+)

 # Configure logging from HINDSIGHT_API_LOG_LEVEL environment variable
 _log_level_str = os.environ.get("HINDSIGHT_API_LOG_LEVEL", "info").lower()
@@ -29,61 +69,213 @@ 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
+        Configured FastMCP server instance
    """
-    # Use stateless_http=True for Claude Code compatibility
-    mcp = FastMCP("hindsight-mcp-server", stateless_http=True)
+    mcp = FastMCP("hindsight-mcp-server")
+
+    global_config = _get_raw_config()
+
+    # Tools available for this mode (multi-bank exposes all tools; single-bank excludes bank-management tools)
+    _SINGLE_BANK_TOOLS: frozenset[str] = frozenset(
+        {
+            "retain",
+            "recall",
+            "reflect",
+            "list_mental_models",
+            "get_mental_model",
+            "create_mental_model",
+            "update_mental_model",
+            "delete_mental_model",
+            "refresh_mental_model",
+            "list_directives",
+            "create_directive",
+            "delete_directive",
+            "list_memories",
+            "get_memory",
+            "delete_memory",
+            "list_documents",
+            "get_document",
+            "delete_document",
+            "list_operations",
+            "get_operation",
+            "cancel_operation",
+            "list_tags",
+            "get_bank",
+            "update_bank",
+            "delete_bank",
+            "clear_memories",
+        }
+    )
+    base_tools: frozenset[str] | None = None if multi_bank else _SINGLE_BANK_TOOLS
+
+    # Apply global mcp_enabled_tools filter (env-level allowlist)
+    if global_config.mcp_enabled_tools is not None:
+        allowed = frozenset(global_config.mcp_enabled_tools)
+        base_tools = (base_tools if base_tools is not None else _ALL_TOOLS) & allowed

    # Configure and register tools using shared module
    config = MCPToolsConfig(
        bank_id_resolver=get_current_bank_id,
-        include_bank_id_param=True,  # HTTP MCP supports multi-bank via parameter
-        tools=None,  # All tools
-        retain_fire_and_forget=False,  # HTTP MCP supports sync/async modes
+        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=base_tools,
    )

    register_mcp_tools(mcp, memory, config)

+    # 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)
+
+    # 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="/", stateless_http=True)
+            self.single_bank_server = create_mcp_server(memory, multi_bank=False)
+            self.single_bank_app = self.single_bank_server.http_app(path="/", stateless_http=True)

    def _get_header(self, scope: dict, name: str) -> str | None:
        """Extract a header value from ASGI scope."""
@@ -95,54 +287,113 @@ 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), extra_headers=e.headers)
+                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
@@ -150,18 +401,29 @@ 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):
+    async def _send_error(self, send, status: int, message: str, extra_headers: dict[str, str] | None = None):
        """Send an error response."""
        body = json.dumps({"error": message}).encode()
+        headers = [(b"content-type", b"application/json")]
+        for key, value in (extra_headers or {}).items():
+            headers.append((key.encode(), value.encode()))
        await send(
            {
                "type": "http.response.start",
                "status": status,
-                "headers": [(b"content-type", b"application/json")],
+                "headers": headers,
            }
        )
        await send(
@@ -172,19 +434,19 @@ 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="/", stateless_http=True)
+
+    single_bank_server = create_mcp_server(memory, multi_bank=False)
+    single_bank_app = single_bank_server.http_app(path="/", stateless_http=True)
+
+    return multi_bank_server, single_bank_server, multi_bank_app, single_bank_app
@@ -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()
@@ -0,0 +1,275 @@
+"""
+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.engine.memory_engine import fq_table
+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(
+                    f"""
+                    SELECT config FROM {fq_table("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(
+                f"""
+                UPDATE {fq_table("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(
+                f"""
+                UPDATE {fq_table("banks")}
+                SET config = '{{}}'::jsonb,
+                    updated_at = now()
+                WHERE bank_id = $1
+                """,
+                bank_id,
+            )
+
+        logger.info(f"Reset bank config for {bank_id} to defaults")
@@ -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
@@ -1,91 +1,83 @@
 """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.
+# Default mission when no bank-specific mission is set
+_DEFAULT_MISSION = "Track every detail: names, numbers, dates, places, and relationships. Prefer specifics over abstractions, never generalise."

-You must output ONLY valid JSON with no markdown formatting, no code blocks, and no additional text.
+# Processing rules — always present regardless of mission
+_PROCESSING_RULES = """Processing rules (always apply):
+- REDUNDANT: same info worded differently → UPDATE the existing observation.
+- CONTRADICTION/UPDATE: capture both states with temporal markers ("used to X, now Y").
+- RESOLVE REFERENCES: when a new fact provides a concrete value resolving a vague placeholder in an existing observation (e.g. "home country", "hometown", "birthplace", "native language", "her ex", "that city"), UPDATE the observation to embed the resolved value explicitly. Example: new fact says "grandma in Sweden" + existing observation says "moved from her home country" → update to "home country is Sweden".
+- NEVER merge observations about different people or unrelated topics."""

-## EXTRACT DURABLE KNOWLEDGE, NOT EPHEMERAL STATE
-Facts often describe events or actions. Extract the DURABLE KNOWLEDGE implied by the fact, not the transient state.
+# Data section — format placeholders {facts_text} and {observations_text} are substituted at call time
+_BATCH_DATA_SECTION = """
+NEW FACTS:
+{facts_text}

-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 history (e.g., "used to X, now Y")
-3. UPDATE: New state replacing old state → update with history
-
-## TAG ROUTING RULES:
-Tags define visibility scopes. The fact and each observation have tags (can be empty = global).
-
-| Fact Tags | Obs Tags | Action |
-|-----------|----------|--------|
-| [alice] | [alice] | UPDATE the observation (same scope) |
-| [alice] | [] | UPDATE the observation (global absorbs all scopes) |
-| [alice] | [bob] | CREATE new untagged observation (cross-scope insight) |
-| [] | [alice] | UPDATE the observation (untagged facts can update any scope) |
-| [] | [] | UPDATE the observation (global to global) |
-
-When NO existing observation matches the fact's topic: CREATE new observation with fact's tags.
-
-## MULTIPLE ACTIONS:
-One fact can trigger MULTIPLE actions. For example:
- Update a scoped observation [alice] about pizza preferences
- AND update a global observation [] about pizza in general
-
-Output an ARRAY of actions (can be empty, one, or many).
-
-## CRITICAL RULES:
- NEVER merge facts about DIFFERENT people
- NEVER merge unrelated topics (food preferences vs work vs hobbies)
- When merging contradictions, capture the CHANGE (before → after)
- Keep observations focused on ONE specific topic per person
- Cross-scope insights (alice's fact about bob's topic) become UNTAGGED (global)
- The "text" field MUST contain durable knowledge, not ephemeral state"""
-
-CONSOLIDATION_USER_PROMPT = """Analyze this new fact and consolidate into knowledge.
-{mission_section}
-NEW FACT: {fact_text}
-FACT TAGS: {fact_tags}
-
-EXISTING OBSERVATIONS:
+EXISTING OBSERVATIONS (JSON array, pooled from recalls across all facts above):
 {observations_text}

-Instructions:
-1. First, extract the DURABLE KNOWLEDGE from the fact (not ephemeral state like "user is at X")
-2. Then compare with existing observations:
-   - If an observation covers the same topic: UPDATE it with the new knowledge
-   - If no observation covers the topic: CREATE a new one
-   - If fact is about different scope: apply tag routing rules
+Each observation includes:
+- id: unique identifier for updating
+- text: the observation content
+- proof_count: number of supporting memories
+- occurred_start/occurred_end: temporal range of source facts
+- source_memories: array of supporting facts with their text and dates

-Output JSON array of actions (ALWAYS an array, even for single action):
-[
-  {{"action": "update", "learning_id": "uuid", "text": "updated durable knowledge", "reason": "..."}},
-  {{"action": "create", "tags": ["tag"], "text": "new durable knowledge", "reason": "..."}}
-]
+Compare the facts against existing observations:
+- Same topic as an existing observation → UPDATE it (observation_id + source_fact_ids)
+- New topic with durable knowledge → CREATE a new observation (source_fact_ids)
+- Cross-reference facts within the batch: a later fact may resolve a vague reference in an earlier one
+- Purely ephemeral facts → omit them (no create/update needed)"""

-If NO consolidation is needed (fact is purely ephemeral with no durable knowledge):
-[]
+# Output format — JSON braces escaped as {{ }} so .format() leaves them literal
+_BATCH_OUTPUT_FORMAT = """
+Output a JSON object with three arrays.

-If no observations exist and fact contains durable knowledge:
-[{{"action": "create", "tags": {fact_tags}, "text": "durable knowledge text", "reason": "new topic"}}]"""
+## EXAMPLE
+
+Input facts:
+[a1b2c3d4-e5f6-7890-abcd-ef1234567890] Alice mentioned she works long hours, often past midnight | Involving: Alice (occurred_start=2024-01-15, mentioned_at=2024-01-15)
+[b2c3d4e5-f6a7-8901-bcde-f12345678901] Alice said she's exhausted from the project deadlines | Involving: Alice (occurred_start=2024-01-20, mentioned_at=2024-01-20)
+
+Good observation text — clean prose, no metadata, each fact tracked distinctly:
+  "Alice works long hours, often past midnight."
+  "Alice feels exhausted from project deadlines."
+
+Bad observation text — NEVER do this (verbatim copy of fact text with metadata):
+  "Alice mentioned she works long hours, often past midnight | Involving: Alice (occurred_start=2024-01-15, mentioned_at=2024-01-15)"
+
+Observation text rules:
+- Write clean prose — NEVER copy raw fact lines or their metadata (temporal fields, "Involving:", "When:" labels, UUIDs).
+- Parenthesized metadata like (occurred_start=...) and pipe-separated labels like "| Involving: ..." are fact formatting — strip them entirely from observation text.
+- How many observations to create and how much to aggregate is driven by the MISSION above.
+
+{{"creates": [{{"text": "Alice works long hours, often past midnight.", "source_fact_ids": ["a1b2c3d4-e5f6-7890-abcd-ef1234567890"]}}, {{"text": "Alice feels exhausted from project deadlines.", "source_fact_ids": ["b2c3d4e5-f6a7-8901-bcde-f12345678901"]}}],
+  "updates": [{{"text": "Alice works at Acme Corp as a senior engineer", "observation_id": "c3d4e5f6-a7b8-9012-cdef-123456789012", "source_fact_ids": ["d4e5f6a7-b8c9-0123-defa-234567890123"]}}],
+  "deletes": [{{"observation_id": "e5f6a7b8-c9d0-1234-efab-345678901234"}}]}}
+
+Rules:
+- "source_fact_ids": copy the EXACT UUID strings shown in brackets [uuid] from NEW FACTS — never use integers or positions.
+- "observation_id": copy the EXACT "id" UUID string from EXISTING OBSERVATIONS.
+- One create/update may reference multiple facts when they jointly support the observation.
+- "deletes": only when an observation is directly superseded or contradicted by new facts.
+- Do NOT include "tags" — handled automatically.
+- Return {{"creates": [], "updates": [], "deletes": []}} if nothing durable is found."""
+
+
+def build_batch_consolidation_prompt(observations_mission: str | None = None) -> str:
+    """
+    Build the consolidation prompt for batch mode (multiple facts per LLM call).
+
+    The mission defines *what* to track (customisable per bank).
+    Processing rules and output format are always present regardless of mission.
+    """
+    mission = observations_mission or _DEFAULT_MISSION
+
+    return (
+        "You are a memory consolidation system. Synthesize facts into observations "
+        "and merge with existing observations when appropriate.\n\n"
+        f"## MISSION\n{mission}\n\n"
+        f"{_PROCESSING_RULES}" + _BATCH_DATA_SECTION + _BATCH_OUTPUT_FORMAT
+    )
@@ -9,6 +9,7 @@ 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

@@ -20,25 +21,29 @@ from ..config import (
    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_LITELLM_API_BASE,
-    ENV_LITELLM_API_KEY,
-    ENV_RERANKER_COHERE_BASE_URL,
+    DEFAULT_RERANKER_ZEROENTROPY_MODEL,
+    ENV_RERANKER_COHERE_API_KEY,
    ENV_RERANKER_COHERE_MODEL,
    ENV_RERANKER_FLASHRANK_CACHE_DIR,
    ENV_RERANKER_FLASHRANK_MODEL,
-    ENV_RERANKER_LITELLM_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,
    ENV_RERANKER_TEI_URL,
+    ENV_RERANKER_ZEROENTROPY_API_KEY,
 )

 logger = logging.getLogger(__name__)
@@ -99,7 +104,13 @@ class LocalSTCrossEncoder(CrossEncoderModel):
    _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):
+    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.

@@ -108,8 +119,15 @@ class LocalSTCrossEncoder(CrossEncoderModel):
                       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

@@ -139,19 +157,47 @@ class LocalSTCrossEncoder(CrossEncoderModel):
        # after loading, which conflicts with accelerate's device_map handling.
        import torch

-        # Check for GPU (CUDA) or Apple Silicon (MPS)
-        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
-        else:
+        # 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}")

-        self._model = CrossEncoder(
-            self.model_name,
-            device=device,
-            model_kwargs={"low_cpu_mem_usage": False},
-        )
+        # 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:
@@ -163,6 +209,11 @@ class LocalSTCrossEncoder(CrossEncoderModel):
        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.
@@ -180,11 +231,11 @@ class LocalSTCrossEncoder(CrossEncoderModel):

        # Use dedicated executor - limited workers naturally limits concurrency
        loop = asyncio.get_event_loop()
-        scores = await loop.run_in_executor(
+        return await loop.run_in_executor(
            LocalSTCrossEncoder._executor,
-            lambda: self._model.predict(pairs, show_progress_bar=False),
+            self._predict_sync,
+            pairs,
        )
-        return scores.tolist() if hasattr(scores, "tolist") else list(scores)


 class RemoteTEICrossEncoder(CrossEncoderModel):
@@ -507,6 +558,104 @@ class CohereCrossEncoder(CrossEncoderModel):
        return all_scores


+class ZeroEntropyCrossEncoder(CrossEncoderModel):
+    """
+    ZeroEntropy cross-encoder implementation using the ZeroEntropy Rerank API.
+
+    Supports zerank-2 (flagship) and zerank-2-small models.
+    See: https://docs.zeroentropy.dev/models
+    """
+
+    RERANK_URL = "https://api.zeroentropy.dev/v1/models/rerank"
+
+    def __init__(
+        self,
+        api_key: str,
+        model: str = DEFAULT_RERANKER_ZEROENTROPY_MODEL,
+        timeout: float = 60.0,
+    ):
+        """
+        Initialize ZeroEntropy cross-encoder client.
+
+        Args:
+            api_key: ZeroEntropy API key
+            model: ZeroEntropy rerank model name (default: zerank-2)
+            timeout: Request timeout in seconds (default: 60.0)
+        """
+        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 "zeroentropy"
+
+    async def initialize(self) -> None:
+        """Initialize the async HTTP client."""
+        if self._async_client is not None:
+            return
+
+        logger.info(f"Reranker: initializing ZeroEntropy provider with model {self.model}")
+        self._async_client = httpx.AsyncClient(
+            timeout=self.timeout,
+            headers={
+                "Authorization": f"Bearer {self.api_key}",
+                "Content-Type": "application/json",
+            },
+        )
+        logger.info("Reranker: ZeroEntropy provider initialized")
+
+    async def predict(self, pairs: list[tuple[str, str]]) -> list[float]:
+        """
+        Score query-document pairs using the ZeroEntropy Rerank API.
+
+        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 for efficient batching
+        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]
+
+            response = await self._async_client.post(
+                self.RERANK_URL,
+                json={
+                    "model": self.model,
+                    "query": query,
+                    "documents": texts,
+                    "top_n": len(texts),
+                },
+            )
+            response.raise_for_status()
+            result = response.json()
+
+            # Map scores back to original positions
+            for item in result.get("results", []):
+                original_idx = item["index"]
+                score = item["relevance_score"]
+                all_scores[indices[original_idx]] = score
+
+        return all_scores
+
+
 class RRFPassthroughCrossEncoder(CrossEncoderModel):
    """
    Passthrough cross-encoder that preserves RRF scores without neural reranking.
@@ -594,7 +743,7 @@ class FlashRankCrossEncoder(CrossEncoderModel):
            return

        try:
-            from flashrank import Ranker  # type: ignore[import-untyped]
+            from flashrank import Ranker
        except ImportError:
            raise ImportError("flashrank is required for FlashRankCrossEncoder. Install it with: pip install flashrank")

@@ -621,7 +770,7 @@ class FlashRankCrossEncoder(CrossEncoderModel):

    def _predict_sync(self, pairs: list[tuple[str, str]]) -> list[float]:
        """Synchronous predict - processes each query group."""
-        from flashrank import RerankRequest  # type: ignore[import-untyped]
+        from flashrank import RerankRequest

        if not pairs:
            return []
@@ -781,50 +930,199 @@ class LiteLLMCrossEncoder(CrossEncoderModel):
        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
-        max_concurrent = int(
-            os.environ.get(ENV_RERANKER_LOCAL_MAX_CONCURRENT, str(DEFAULT_RERANKER_LOCAL_MAX_CONCURRENT))
+        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,
        )
-        return LocalSTCrossEncoder(model_name=model_name, max_concurrent=max_concurrent)
    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)
-        base_url = os.environ.get(ENV_RERANKER_COHERE_BASE_URL) or None
-        return CohereCrossEncoder(api_key=api_key, model=model, base_url=base_url)
+            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":
-        api_base = os.environ.get(ENV_LITELLM_API_BASE, DEFAULT_LITELLM_API_BASE)
-        api_key = os.environ.get(ENV_LITELLM_API_KEY)
-        model = os.environ.get(ENV_RERANKER_LITELLM_MODEL, DEFAULT_RERANKER_LITELLM_MODEL)
-        return LiteLLMCrossEncoder(api_base=api_base, api_key=api_key, model=model)
+        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 == "zeroentropy":
+        api_key = config.reranker_zeroentropy_api_key
+        if not api_key:
+            raise ValueError(
+                f"{ENV_RERANKER_ZEROENTROPY_API_KEY} is required when {ENV_RERANKER_PROVIDER} is 'zeroentropy'"
+            )
+        return ZeroEntropyCrossEncoder(
+            api_key=api_key,
+            model=config.reranker_zeroentropy_model,
+        )
    elif provider == "rrf":
        return RRFPassthroughCrossEncoder()
    else:
        raise ValueError(
-            f"Unknown reranker provider: {provider}. Supported: 'local', 'tei', 'cohere', 'flashrank', 'litellm', 'rrf'"
+            f"Unknown reranker provider: {provider}. Supported: 'local', 'tei', 'cohere', 'zeroentropy', 'flashrank', 'litellm', 'litellm-sdk', 'rrf'"
        )
@@ -20,6 +20,7 @@ RETRYABLE_EXCEPTIONS = (
    asyncpg.exceptions.InterfaceError,
    asyncpg.exceptions.ConnectionDoesNotExistError,
    asyncpg.exceptions.TooManyConnectionsError,
+    asyncpg.exceptions.DeadlockDetectedError,
    OSError,
    ConnectionError,
    asyncio.TimeoutError,
@@ -11,6 +11,7 @@ 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
@@ -18,22 +19,23 @@ 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,
    DEFAULT_LITELLM_API_BASE,
-    ENV_COHERE_API_KEY,
-    ENV_EMBEDDINGS_COHERE_BASE_URL,
-    ENV_EMBEDDINGS_COHERE_MODEL,
-    ENV_EMBEDDINGS_LITELLM_MODEL,
+    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,
-    ENV_LITELLM_API_BASE,
-    ENV_LITELLM_API_KEY,
    ENV_LLM_API_KEY,
 )

@@ -92,15 +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

@@ -134,19 +143,47 @@ class LocalSTEmbeddings(Embeddings):
        # which can cause issues when accelerate is installed but no GPU is available.
        import torch

-        # Check for GPU (CUDA) or Apple Silicon (MPS)
-        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
-        else:
+        # 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}")

-        self._model = SentenceTransformer(
-            self.model_name,
-            device=device,
-            model_kwargs={"low_cpu_mem_usage": False},
-        )
+        # 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})")
@@ -163,6 +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]

@@ -529,7 +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})")
@@ -684,26 +722,175 @@ class LiteLLMEmbeddings(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,
+                "encoding_format": "float",
+            }
+            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,
+                    "encoding_format": "float",
+                }
+                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)
@@ -716,18 +903,33 @@ def create_embeddings_from_env() -> Embeddings:
        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)
-        base_url = os.environ.get(ENV_EMBEDDINGS_COHERE_BASE_URL) or None
-        return CohereEmbeddings(api_key=api_key, model=model, base_url=base_url)
+            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":
-        api_base = os.environ.get(ENV_LITELLM_API_BASE, DEFAULT_LITELLM_API_BASE)
-        api_key = os.environ.get(ENV_LITELLM_API_KEY)
-        model = os.environ.get(ENV_EMBEDDINGS_LITELLM_MODEL, DEFAULT_EMBEDDINGS_LITELLM_MODEL)
-        return LiteLLMEmbeddings(api_base=api_base, api_key=api_key, model=model)
+        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', 'litellm'"
+            f"Unknown embeddings provider: {provider}. "
+            f"Supported: 'local', 'tei', 'openai', 'cohere', 'litellm', 'litellm-sdk'"
        )
@@ -5,6 +5,10 @@ Uses spaCy for entity extraction and implements resolution logic
 to disambiguate entities across memory units.
 """

+import asyncio
+import logging
+from collections import defaultdict
+from dataclasses import dataclass, field
 from datetime import UTC, datetime
 from difflib import SequenceMatcher

@@ -12,6 +16,43 @@ import asyncpg

 from .db_utils import acquire_with_retry
 from .memory_engine import fq_table
+from .retain.entity_labels import build_labels_lookup as _build_labels_lookup_from_config
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class _EntityToCreate:
+    """An entity that needs to be inserted (no matching candidate found)."""
+
+    idx: int
+    name: str
+    event_date: datetime | None
+
+
+@dataclass
+class _EntityStat:
+    """Stat accumulation entry for a resolved entity (post-transaction update)."""
+
+    entity_id: str
+    event_date: datetime | None
+
+
+@dataclass
+class _EntityStatAgg:
+    """Aggregated stats used when flushing pending updates."""
+
+    count: int = 0
+    max_date: datetime | None = None
+
+
+@dataclass
+class _CooccurrencePair:
+    """A (entity_id_1, entity_id_2) pair observed in a retain batch (for post-txn flush)."""
+
+    entity_id_1: str
+    entity_id_2: str
+

 # Load spaCy model (singleton)
 _nlp = None
@@ -22,14 +63,95 @@ class EntityResolver:
    Resolves entities to canonical IDs with disambiguation.
    """

-    def __init__(self, pool: asyncpg.Pool):
+    def __init__(self, pool: asyncpg.Pool, entity_lookup: str = "full"):
        """
        Initialize entity resolver.

        Args:
            pool: asyncpg connection pool
+            entity_lookup: Lookup strategy — "full" loads all bank entities then
+                matches in Python; "trigram" uses pg_trgm GIN index to fetch only
+                similar candidates per entity name (much faster for large banks).
        """
        self.pool = pool
+        self.entity_lookup = entity_lookup
+        # Keyed by asyncio task id so concurrent retain batches never mix their
+        # pending updates.  flush_pending_stats() pops only the calling task's items.
+        self._pending_stats: dict[int, list[_EntityStat]] = {}
+        self._pending_cooccurrences: dict[int, list[_CooccurrencePair]] = {}
+
+    def _task_key(self) -> int:
+        """Return a unique key for the current asyncio task (or 0 for non-task context)."""
+        task = asyncio.current_task()
+        return id(task) if task is not None else 0
+
+    async def flush_pending_stats(self) -> None:
+        """
+        Flush accumulated entity stats and co-occurrence counts for the current task.
+
+        Must be called AFTER the retain transaction commits.  Pops only the items
+        accumulated by the calling asyncio task so concurrent retain batches never
+        flush each other's uncommitted entity IDs.
+        """
+        if self.pool is None:
+            return
+
+        key = self._task_key()
+        stats = self._pending_stats.pop(key, [])
+        cooccurrences = self._pending_cooccurrences.pop(key, [])
+
+        if not stats and not cooccurrences:
+            return
+
+        async with acquire_with_retry(self.pool) as conn:
+            if stats:
+                # Aggregate: sum counts and find max date per entity_id.
+                agg: dict[str, _EntityStatAgg] = defaultdict(_EntityStatAgg)
+                for s in stats:
+                    entry = agg[s.entity_id]
+                    entry.count += 1
+                    if s.event_date is not None:
+                        entry.max_date = s.event_date if entry.max_date is None else max(entry.max_date, s.event_date)
+
+                # Sort by entity_id so all concurrent workers acquire row locks in
+                # the same order — prevents circular lock dependencies (deadlocks).
+                rows = sorted((eid, a.count, a.max_date) for eid, a in agg.items())
+                await conn.executemany(
+                    f"""
+                    UPDATE {fq_table("entities")} SET
+                        mention_count = mention_count + $2,
+                        last_seen     = GREATEST(last_seen, $3)
+                    WHERE id = $1::uuid
+                    """,
+                    rows,
+                )
+
+            if cooccurrences:
+                # Aggregate: count occurrences per (entity_id_1, entity_id_2) pair.
+                coo_agg: dict[tuple[str, str], int] = {}
+                for c in cooccurrences:
+                    pair = (c.entity_id_1, c.entity_id_2)
+                    coo_agg[pair] = coo_agg.get(pair, 0) + 1
+
+                now = datetime.now(UTC)
+                # Sort by (entity_id_1, entity_id_2) for consistent lock ordering.
+                await conn.executemany(
+                    f"""
+                    INSERT INTO {fq_table("entity_cooccurrences")}
+                        (entity_id_1, entity_id_2, cooccurrence_count, last_cooccurred)
+                    VALUES ($1, $2, $3, $4)
+                    ON CONFLICT (entity_id_1, entity_id_2)
+                    DO UPDATE SET
+                        cooccurrence_count = {fq_table("entity_cooccurrences")}.cooccurrence_count + EXCLUDED.cooccurrence_count,
+                        last_cooccurred    = GREATEST({fq_table("entity_cooccurrences")}.last_cooccurred, EXCLUDED.last_cooccurred)
+                    """,
+                    sorted((e1, e2, count, now) for (e1, e2), count in coo_agg.items()),
+                )
+
+    @staticmethod
+    def _build_labels_lookup(entity_labels: list | None) -> set[str]:
+        """Build a set of valid 'key:value' entity label strings for fast lookup."""
+        return _build_labels_lookup_from_config(entity_labels)

    async def resolve_entities_batch(
        self,
@@ -38,6 +160,7 @@ class EntityResolver:
        context: str,
        unit_event_date,
        conn=None,
+        entity_labels: list | None = None,
    ) -> list[str]:
        """
        Resolve multiple entities in batch (MUCH faster than sequential).
@@ -58,15 +181,34 @@ class EntityResolver:
        if not entities_data:
            return []

+        taxonomy_lookup = self._build_labels_lookup(entity_labels)
        if conn is None:
            async with acquire_with_retry(self.pool) as conn:
-                return await self._resolve_entities_batch_impl(conn, bank_id, entities_data, context, unit_event_date)
+                return await self._resolve_entities_batch_impl(
+                    conn, bank_id, entities_data, context, unit_event_date, taxonomy_lookup
+                )
        else:
-            return await self._resolve_entities_batch_impl(conn, bank_id, entities_data, context, unit_event_date)
+            return await self._resolve_entities_batch_impl(
+                conn, bank_id, entities_data, context, unit_event_date, taxonomy_lookup
+            )

    async def _resolve_entities_batch_impl(
-        self, conn, bank_id: str, entities_data: list[dict], context: str, unit_event_date
+        self,
+        conn,
+        bank_id: str,
+        entities_data: list[dict],
+        context: str,
+        unit_event_date,
+        taxonomy_lookup: set[str] | None = None,
    ) -> list[str]:
+        if self.entity_lookup == "trigram":
+            return await self._resolve_entities_batch_trigram(conn, bank_id, entities_data, unit_event_date)
+        return await self._resolve_entities_batch_full(conn, bank_id, entities_data, unit_event_date)
+
+    async def _resolve_entities_batch_full(
+        self, conn, bank_id: str, entities_data: list[dict], unit_event_date
+    ) -> list[str]:
+        """Original strategy: load all bank entities then match in Python."""
        # Query ALL candidates for this bank
        all_entities = await conn.fetch(
            f"""
@@ -130,10 +272,103 @@ class EntityResolver:
                    matching.append((ent_id, canonical_name, metadata, last_seen, mention_count))
            all_candidates[entity_text] = matching

+        return await self._resolve_from_candidates(
+            conn, bank_id, entities_data, unit_event_date, all_candidates, cooccurrence_map
+        )
+
+    async def _resolve_entities_batch_trigram(
+        self, conn, bank_id: str, entities_data: list[dict], unit_event_date
+    ) -> list[str]:
+        """
+        Trigram strategy: fetch only similar candidates per entity name using pg_trgm.
+
+        Instead of loading all bank entities (O(N)), uses a GIN trigram index to fetch
+        only the small set of candidates that are textually similar to each input name.
+        Reduces DB data transfer from 165K rows to ~5-20 rows per entity.
+        """
+        entity_texts = list(set(e["text"] for e in entities_data))
+
+        # Fetch candidates for all unique entity texts in a single batched query.
+        # The trigram % operator uses the GIN index; the substring conditions cover
+        # exact prefix/suffix matches that trigrams might miss at low similarity.
+        rows = await conn.fetch(
+            f"""
+            SELECT DISTINCT ON (e.id)
+                e.id, e.canonical_name, e.metadata, e.last_seen, e.mention_count,
+                q.query_text
+            FROM unnest($2::text[]) AS q(query_text)
+            JOIN {fq_table("entities")} e ON (
+                e.bank_id = $1
+                AND (
+                    e.canonical_name % q.query_text
+                    OR LOWER(e.canonical_name) LIKE '%' || LOWER(q.query_text) || '%'
+                    OR LOWER(q.query_text) LIKE '%' || LOWER(e.canonical_name) || '%'
+                )
+            )
+            """,
+            bank_id,
+            entity_texts,
+        )
+
+        # Group candidates by query_text
+        all_candidates: dict[str, list] = {t: [] for t in entity_texts}
+        candidate_ids: set = set()
+        for row in rows:
+            query_text = row["query_text"]
+            all_candidates[query_text].append(
+                (row["id"], row["canonical_name"], row["metadata"], row["last_seen"], row["mention_count"])
+            )
+            candidate_ids.add(row["id"])
+
+        # Fetch co-occurrences only for the candidate entities (not all bank entities)
+        cooccurrence_map: dict[str, set[str]] = {}
+        if candidate_ids:
+            candidate_id_list = list(candidate_ids)
+            cooc_rows = await conn.fetch(
+                f"""
+                SELECT ec.entity_id_1, ec.entity_id_2
+                FROM {fq_table("entity_cooccurrences")} ec
+                WHERE ec.entity_id_1 = ANY($1::uuid[])
+                   OR ec.entity_id_2 = ANY($1::uuid[])
+                """,
+                candidate_id_list,
+            )
+            # Build name lookup for co-occurrence mapping
+            id_to_name = {
+                row["id"]: row["canonical_name"].lower()
+                for cands in all_candidates.values()
+                for row in [{"id": c[0], "canonical_name": c[1]} for c in cands]
+            }
+            for row in cooc_rows:
+                eid1, eid2 = row["entity_id_1"], row["entity_id_2"]
+                if eid1 not in cooccurrence_map:
+                    cooccurrence_map[eid1] = set()
+                if eid2 not in cooccurrence_map:
+                    cooccurrence_map[eid2] = set()
+                if eid2 in id_to_name:
+                    cooccurrence_map[eid1].add(id_to_name[eid2])
+                if eid1 in id_to_name:
+                    cooccurrence_map[eid2].add(id_to_name[eid1])
+
+        return await self._resolve_from_candidates(
+            conn, bank_id, entities_data, unit_event_date, all_candidates, cooccurrence_map
+        )
+
+    async def _resolve_from_candidates(
+        self,
+        conn,
+        bank_id: str,
+        entities_data: list[dict],
+        unit_event_date,
+        all_candidates: dict[str, list],
+        cooccurrence_map: dict[str, set[str]],
+    ) -> list[str]:
+        """Shared scoring + upsert logic used by both lookup strategies."""
+
        # Resolve each entity using pre-fetched candidates
        entity_ids = [None] * len(entities_data)
-        entities_to_update = []  # (entity_id, event_date)
-        entities_to_create = []  # (idx, entity_data, event_date)
+        entities_to_update: list[_EntityStat] = []
+        entities_to_create: list[_EntityToCreate] = []

        for idx, entity_data in enumerate(entities_data):
            entity_text = entity_data["text"]
@@ -145,7 +380,7 @@ class EntityResolver:

            if not candidates:
                # Will create new entity
-                entities_to_create.append((idx, entity_data, entity_event_date))
+                entities_to_create.append(_EntityToCreate(idx=idx, name=entity_text, event_date=entity_event_date))
                continue

            # Score candidates
@@ -189,73 +424,83 @@ class EntityResolver:

            if best_score > threshold:
                entity_ids[idx] = best_candidate
-                entities_to_update.append((best_candidate, entity_event_date))
+                entities_to_update.append(_EntityStat(entity_id=best_candidate, event_date=entity_event_date))
            else:
-                entities_to_create.append((idx, entity_data, entity_event_date))
+                entities_to_create.append(
+                    _EntityToCreate(idx=idx, name=entity_data["text"], event_date=entity_event_date)
+                )

-        # Batch update existing entities
-        if entities_to_update:
-            await conn.executemany(
-                f"""
-                UPDATE {fq_table("entities")} SET
-                    mention_count = mention_count + 1,
-                    last_seen = $2
-                WHERE id = $1::uuid
-                """,
-                entities_to_update,
-            )
+        # Existing entities: IDs already known from the candidate SELECT above.
+        # No in-transaction UPDATE — mention_count/last_seen are stats deferred to
+        # flush_pending_stats() which the orchestrator calls after the transaction.
+        pending: list[_EntityStat] = list(entities_to_update)

-        # Batch create new entities using COPY + INSERT for maximum speed
-        # This handles duplicates via ON CONFLICT and returns all IDs
+        # New entities: INSERT with DO NOTHING to avoid row locks on concurrent races.
+        # ON CONFLICT DO NOTHING returns nothing for rows that conflicted; we handle
+        # that rare case with a fallback SELECT.
        if entities_to_create:
-            # Group entities by canonical name (lowercase) to handle duplicates within batch
-            # For duplicates, we only insert once and reuse the ID, but track the count
-            unique_entities = {}  # lowercase_name -> (entity_data, event_date, [indices])
-            for idx, entity_data, event_date in entities_to_create:
-                name_lower = entity_data["text"].lower()
-                if name_lower not in unique_entities:
-                    unique_entities[name_lower] = (entity_data, event_date, [idx])
-                else:
-                    # Same entity appears multiple times - add index to list
-                    unique_entities[name_lower][2].append(idx)
+            # Group by lowercase name — deduplicate within the batch.
+            @dataclass
+            class _NameGroup:
+                name: str
+                event_date: datetime | None
+                indices: list[int] = field(default_factory=list)

-            # Batch insert unique entities and get their IDs
-            # Use a single query with unnest for speed
-            entity_names = []
-            entity_dates = []
-            entity_counts = []  # Track how many times each entity appears in this batch
-            indices_map = []  # Maps result index -> list of original indices
+            groups: dict[str, _NameGroup] = {}
+            for e in entities_to_create:
+                name_lower = e.name.lower()
+                if name_lower not in groups:
+                    groups[name_lower] = _NameGroup(name=e.name, event_date=e.event_date)
+                groups[name_lower].indices.append(e.idx)

-            for name_lower, (entity_data, event_date, indices) in unique_entities.items():
-                entity_names.append(entity_data["text"])
-                entity_dates.append(event_date)
-                entity_counts.append(len(indices))  # Count of occurrences in this batch
-                indices_map.append(indices)
+            # Sort by lowercase name for deterministic ordering.
+            sorted_groups = sorted(groups.items())
+            entity_names = [g.name for _, g in sorted_groups]
+            entity_dates = [g.event_date for _, g in sorted_groups]

-            # Batch INSERT ... ON CONFLICT with RETURNING
-            # Uses the batch count for mention_count instead of always 1
-            rows = await conn.fetch(
+            # INSERT ... ON CONFLICT DO NOTHING — no row lock on already-existing entities.
+            inserted_rows = await conn.fetch(
                f"""
                INSERT INTO {fq_table("entities")} (bank_id, canonical_name, first_seen, last_seen, mention_count)
-                SELECT $1, name, event_date, event_date, cnt
-                FROM unnest($2::text[], $3::timestamptz[], $4::int[]) AS t(name, event_date, cnt)
+                SELECT $1, name, COALESCE(event_date, now()), COALESCE(event_date, now()), 1
+                FROM unnest($2::text[], $3::timestamptz[]) AS t(name, event_date)
                ON CONFLICT (bank_id, LOWER(canonical_name))
-                DO UPDATE SET
-                    mention_count = {fq_table("entities")}.mention_count + EXCLUDED.mention_count,
-                    last_seen = EXCLUDED.last_seen
-                RETURNING id
+                DO NOTHING
+                RETURNING id, LOWER(canonical_name) AS name_lower
                """,
                bank_id,
                entity_names,
                entity_dates,
-                entity_counts,
            )
+            id_by_name: dict[str, str] = {row["name_lower"]: row["id"] for row in inserted_rows}

-            # Map returned IDs back to original indices
-            for result_idx, row in enumerate(rows):
-                entity_id = row["id"]
-                for original_idx in indices_map[result_idx]:
-                    entity_ids[original_idx] = entity_id
+            # Fallback SELECT for names that conflicted (another worker won the race).
+            missing = [n for n, _ in sorted_groups if n not in id_by_name]
+            if missing:
+                existing_rows = await conn.fetch(
+                    f"""
+                    SELECT id, LOWER(canonical_name) AS name_lower
+                    FROM {fq_table("entities")}
+                    WHERE bank_id = $1 AND LOWER(canonical_name) = ANY($2::text[])
+                    """,
+                    bank_id,
+                    missing,
+                )
+                for row in existing_rows:
+                    id_by_name[row["name_lower"]] = row["id"]
+
+            # Assign entity IDs back and queue for post-txn stats flush.
+            for name_lower, g in sorted_groups:
+                entity_id = id_by_name.get(name_lower)
+                if entity_id:
+                    for original_idx in g.indices:
+                        entity_ids[original_idx] = entity_id
+                    pending.append(_EntityStat(entity_id=entity_id, event_date=g.event_date))
+
+        # Accumulate into the resolver's pending list; the orchestrator flushes
+        # these with await entity_resolver.flush_pending_stats() after the txn.
+        key = self._task_key()
+        self._pending_stats.setdefault(key, []).extend(pending)

        return entity_ids

@@ -408,7 +653,7 @@ class EntityResolver:
        entity_id = await conn.fetchval(
            f"""
            INSERT INTO {fq_table("entities")} (bank_id, canonical_name, first_seen, last_seen, mention_count)
-            VALUES ($1, $2, $3, $4, 1)
+            VALUES ($1, $2, COALESCE($3, now()), COALESCE($4, now()), 1)
            ON CONFLICT (bank_id, LOWER(canonical_name))
            DO UPDATE SET
                mention_count = {fq_table("entities")}.mention_count + 1,
@@ -541,19 +786,14 @@ class EntityResolver:
                        entity_id_1, entity_id_2 = entity_id_2, entity_id_1
                    cooccurrence_pairs.add((entity_id_1, entity_id_2))

-        # Batch update co-occurrences
+        # Accumulate co-occurrence pairs for post-transaction flush.
+        # The actual INSERT/UPDATE is deferred to flush_pending_stats() to avoid
+        # row-level lock contention (ON CONFLICT DO UPDATE inside a long transaction
+        # serialises concurrent writers on popular entity pairs).
        if cooccurrence_pairs:
-            now = datetime.now(UTC)
-            await conn.executemany(
-                f"""
-                INSERT INTO {fq_table("entity_cooccurrences")} (entity_id_1, entity_id_2, cooccurrence_count, last_cooccurred)
-                VALUES ($1, $2, $3, $4)
-                ON CONFLICT (entity_id_1, entity_id_2)
-                DO UPDATE SET
-                    cooccurrence_count = {fq_table("entity_cooccurrences")}.cooccurrence_count + 1,
-                    last_cooccurred = EXCLUDED.last_cooccurred
-                """,
-                [(e1, e2, 1, now) for e1, e2 in cooccurrence_pairs],
+            key = self._task_key()
+            self._pending_cooccurrences.setdefault(key, []).extend(
+                _CooccurrencePair(entity_id_1=e1, entity_id_2=e2) for e1, e2 in cooccurrence_pairs
            )

    async def get_units_by_entity(self, entity_id: str, limit: int = 100) -> list[str]:
@@ -12,6 +12,7 @@ from typing import TYPE_CHECKING, Any
 if TYPE_CHECKING:
    from hindsight_api.engine.memory_engine import Budget
    from hindsight_api.engine.response_models import RecallResult, ReflectResult
+    from hindsight_api.engine.search.tags import TagsMatch
    from hindsight_api.models import RequestContext


@@ -48,6 +49,7 @@ 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 +57,9 @@ 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.
@@ -335,6 +338,8 @@ class MemoryEngineInterface(ABC):
        bank_id: str,
        *,
        search_query: str | None = None,
+        tags: list[str] | None = None,
+        tags_match: "TagsMatch" = "any_strict",
        limit: int = 100,
        offset: int = 0,
        request_context: "RequestContext",
@@ -344,7 +349,9 @@ class MemoryEngineInterface(ABC):

        Args:
            bank_id: The memory bank ID.
-            search_query: Search query.
+            search_query: Case-insensitive substring filter on document ID.
+            tags: Filter by tags.
+            tags_match: How to match tags (any, all, any_strict, all_strict).
            limit: Maximum results.
            offset: Pagination offset.
            request_context: Request context for authentication.
@@ -442,49 +449,6 @@ class MemoryEngineInterface(ABC):
        """
        ...

-    @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.
-        """
-        ...
-
    # =========================================================================
    # Statistics & Operations
    # =========================================================================
@@ -604,6 +568,7 @@ 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.
@@ -612,6 +577,7 @@ 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.
@@ -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
@@ -1,14 +0,0 @@
-"""
-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"]
@@ -1,53 +0,0 @@
-"""
-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"
-    )
@@ -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)
@@ -0,0 +1,128 @@
+"""File parser implementations."""
+
+import logging
+from dataclasses import dataclass
+
+from .base import FileParser, UnsupportedFileTypeError
+from .iris import IrisParser
+from .markitdown import MarkitdownParser
+
+__all__ = [
+    "FileParser",
+    "UnsupportedFileTypeError",
+    "IrisParser",
+    "MarkitdownParser",
+    "FileParserRegistry",
+    "ConvertResult",
+]
+
+
+@dataclass
+class ConvertResult:
+    """Result of a successful file conversion."""
+
+    content: str
+    parser_name: str
+
+
+logger = logging.getLogger(__name__)
+
+
+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())}")
+
+    async def convert_with_fallback(
+        self,
+        parsers: list[str],
+        file_data: bytes,
+        filename: str,
+        content_type: str | None = None,
+    ) -> ConvertResult:
+        """
+        Try each parser in order, falling back on failure or empty content.
+
+        Moves to the next parser if the current one raises UnsupportedFileTypeError
+        or returns empty content. Any other exception (RuntimeError, network error,
+        etc.) also triggers a fallback so the chain is exhausted before failing.
+
+        Args:
+            parsers: Ordered list of parser names to try
+            file_data: Raw file bytes
+            filename: Original filename
+            content_type: MIME type (optional)
+
+        Returns:
+            ConvertResult with the parsed content and the name of the parser that succeeded
+
+        Raises:
+            ValueError: If a parser name is not registered
+            RuntimeError: If all parsers fail or return empty content
+        """
+        last_error: Exception | None = None
+        for name in parsers:
+            parser = self.get_parser(name, filename, content_type)
+            try:
+                content = await parser.convert(file_data, filename)
+                if content and content.strip():
+                    return ConvertResult(content=content, parser_name=name)
+                logger.warning(f"Parser '{name}' returned empty content for '{filename}', trying next")
+                last_error = RuntimeError(f"Parser '{name}' returned no content for '{filename}'")
+            except UnsupportedFileTypeError as e:
+                logger.warning(f"Parser '{name}' does not support '{filename}', trying next: {e}")
+                last_error = e
+            except Exception as e:
+                logger.warning(f"Parser '{name}' failed for '{filename}', trying next: {e}")
+                last_error = e
+
+        raise last_error or RuntimeError(f"No parsers available for '{filename}'")
+
+    def list_parsers(self) -> list[str]:
+        """Get list of registered parser names."""
+        return list(self._parsers.keys())
@@ -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
@@ -0,0 +1,138 @@
+"""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(timeout=httpx.Timeout(30.0, read=120.0)) 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)
+            # Ensure file_data is plain bytes (GCS storage may return obstore.Bytes)
+            upload_resp = await client.put(
+                upload_url,
+                content=bytes(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)
@@ -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"
@@ -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"]
@@ -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()
@@ -0,0 +1,513 @@
+"""
+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
+                try:
+                    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 result.model_dump_json(),
+                        input_tokens=estimated_input,
+                        output_tokens=estimated_output,
+                        duration=duration,
+                        finish_reason=None,
+                        error=None,
+                    )
+                except Exception:
+                    pass  # logging failure must never affect the operation
+
+                # 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
@@ -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()
@@ -0,0 +1,635 @@
+"""
+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 contextvars import ContextVar
+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.llm_wrapper import parse_llm_json
+from hindsight_api.engine.response_models import LLMToolCall, LLMToolCallResult, TokenUsage
+from hindsight_api.metrics import get_metrics_collector
+
+logger = logging.getLogger(__name__)
+
+# Per-request Gemini safety settings override.
+# Set exclusively by ConfiguredLLMProvider.call() / call_with_tools() via token-based
+# set/reset, so it is properly scoped to each individual LLM call and never leaks.
+_safety_settings_ctx: ContextVar[list | None] = ContextVar("gemini_safety_settings", default=None)
+
+
+# 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"
+
+        # Safety settings: None means use Gemini's defaults
+        self._safety_settings: list | None = kwargs.get("gemini_safety_settings")
+
+        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
+
+        # Apply safety settings: context var (per-request bank override) takes precedence over instance default
+        effective_safety_settings = _safety_settings_ctx.get()
+        if effective_safety_settings is None:
+            effective_safety_settings = self._safety_settings
+        if effective_safety_settings is not None:
+            config_kwargs["safety_settings"] = [
+                genai_types.SafetySetting(category=s["category"], threshold=s["threshold"])
+                for s in effective_safety_settings
+            ]
+
+        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 asyncio.wait_for(
+                    self._client.aio.models.generate_content(
+                        model=self.model,
+                        contents=gemini_contents,
+                        config=generation_config,
+                    ),
+                    timeout=90.0,  # Safety net for network hangs; valid slow responses are <90s
+                )
+
+                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 = parse_llm_json(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 = []
+        msg_list = list(messages)
+        i = 0
+        while i < len(msg_list):
+            msg = msg_list[i]
+            role = msg.get("role", "user")
+            content = msg.get("content", "")
+
+            if role == "system":
+                system_instruction = (system_instruction + "\n\n" + content) if system_instruction else content
+                i += 1
+            elif role == "tool":
+                # Gemini requires ALL tool responses for a given model turn to be grouped
+                # into a single Content with multiple FunctionResponse parts.
+                # Consecutive role="tool" messages correspond to one model turn's tool calls.
+                parts = []
+                while i < len(msg_list) and msg_list[i].get("role") == "tool":
+                    tool_msg = msg_list[i]
+                    tool_content = tool_msg.get("content", "")
+                    parts.append(
+                        genai_types.Part(
+                            function_response=genai_types.FunctionResponse(
+                                name=tool_msg.get("name", ""),
+                                response={"result": tool_content},
+                            )
+                        )
+                    )
+                    i += 1
+                gemini_contents.append(genai_types.Content(role="user", parts=parts))
+            elif role == "assistant":
+                tool_calls_in_msg = msg.get("tool_calls", [])
+                if tool_calls_in_msg:
+                    # Convert OpenAI-style tool_calls to Gemini function_call parts
+                    # This is required for proper multi-turn conversation history
+                    parts = []
+                    if content:
+                        parts.append(genai_types.Part(text=content))
+                    for tc in tool_calls_in_msg:
+                        fn = tc.get("function", {})
+                        fn_name = fn.get("name", "")
+                        fn_args_str = fn.get("arguments", "{}")
+                        fn_args = parse_llm_json(fn_args_str)
+                        parts.append(
+                            genai_types.Part(function_call=genai_types.FunctionCall(name=fn_name, args=fn_args))
+                        )
+                    gemini_contents.append(genai_types.Content(role="model", parts=parts))
+                else:
+                    gemini_contents.append(genai_types.Content(role="model", parts=[genai_types.Part(text=content)]))
+                i += 1
+            else:
+                gemini_contents.append(genai_types.Content(role="user", parts=[genai_types.Part(text=content)]))
+                i += 1
+
+        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
+
+        # Map OpenAI-style tool_choice to Gemini FunctionCallingConfig
+        if tool_choice == "required":
+            config_kwargs["tool_config"] = genai_types.ToolConfig(
+                function_calling_config=genai_types.FunctionCallingConfig(
+                    mode="ANY",
+                )
+            )
+        elif isinstance(tool_choice, dict) and tool_choice.get("type") == "function":
+            fn_name = tool_choice.get("function", {}).get("name")
+            if fn_name:
+                config_kwargs["tool_config"] = genai_types.ToolConfig(
+                    function_calling_config=genai_types.FunctionCallingConfig(
+                        mode="ANY",
+                        allowed_function_names=[fn_name],
+                    )
+                )
+        elif tool_choice == "none":
+            config_kwargs["tool_config"] = genai_types.ToolConfig(
+                function_calling_config=genai_types.FunctionCallingConfig(mode="NONE")
+            )
+        # "auto" is the default (no tool_config needed)
+
+        # Apply safety settings: context var (per-request bank override) takes precedence over instance default
+        effective_safety_settings = _safety_settings_ctx.get()
+        if effective_safety_settings is None:
+            effective_safety_settings = self._safety_settings
+        if effective_safety_settings is not None:
+            config_kwargs["safety_settings"] = [
+                genai_types.SafetySetting(category=s["category"], threshold=s["threshold"])
+                for s in effective_safety_settings
+            ]
+
+        config = genai_types.GenerateContentConfig(**config_kwargs)
+
+        last_exception = None
+        for attempt in range(max_retries + 1):
+            try:
+                response = await asyncio.wait_for(
+                    self._client.aio.models.generate_content(
+                        model=self.model,
+                        contents=gemini_contents,
+                        config=config,
+                    ),
+                    timeout=90.0,  # Safety net for network hangs; valid slow responses are <90s
+                )
+
+                # 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
@@ -0,0 +1,323 @@
+"""
+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 collections.abc import Callable
+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
+        self._response_callback: Callable[[list[dict], str], Any] | 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._response_callback is not None:
+            result = self._response_callback(messages, scope)
+        elif 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._response_callback is not None:
+            cb_result = self._response_callback(messages, scope)
+            if isinstance(cb_result, LLMToolCallResult):
+                result = cb_result
+            else:
+                result = LLMToolCallResult(
+                    content=str(cb_result) if cb_result is not None else "mock response", finish_reason="stop"
+                )
+        elif 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_response_callback(self, fn: Callable[[list[dict], str], Any]) -> None:
+        """
+        Set a callback invoked on each call() instead of _mock_response.
+
+        The callback receives (messages, scope) and returns the response.
+        Useful for returning different responses per call (e.g., cycling
+        through a corpus in a benchmark).
+        """
+        self._response_callback = fn
+
+    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
@@ -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()
@@ -92,11 +92,17 @@ class DateparserQueryAnalyzer(QueryAnalyzer):
        self._search_dates = None

    def load(self) -> None:
-        """Load dateparser (lazy import)."""
+        """Load dateparser and warm up internal data structures.
+
+        Triggers the real initialization cost (regex tables, timezone data) at
+        load time so the first actual recall doesn't pay the cold-start penalty.
+        """
        if self._search_dates is None:
            from dateparser.search import search_dates

            self._search_dates = search_dates
+            # Warm up: fire a dummy call to trigger lazy-loaded internal tables.
+            self._search_dates("today")

    def analyze(self, query: str, reference_date: datetime | None = None) -> QueryAnalysis:
        """
@@ -14,32 +14,26 @@ import re
 import time
 from typing import TYPE_CHECKING, Any, Awaitable, Callable

+import tiktoken
+
 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
-    """
+    """Build list of DirectiveInfo from directives."""
    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
+    return [
+        DirectiveInfo(
+            id=directive.get("id", ""),
+            name=directive.get("name", ""),
+            content=directive.get("content", ""),
+        )
+        for directive in directives
+    ]


 if TYPE_CHECKING:
@@ -58,6 +52,7 @@ def _normalize_tool_name(name: str) -> str:
    - '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.)
    """
@@ -69,6 +64,11 @@ def _normalize_tool_name(name: str) -> str:
    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


@@ -80,6 +80,18 @@ def _is_done_tool(name: str) -> bool:
 # 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.
@@ -92,6 +104,33 @@ def _clean_answer_text(text: str) -> str:
    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,
@@ -141,35 +180,55 @@ async def _generate_structured_output(
            fields[field_name] = (field_type, default)

        if not fields:
-            return None
+            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"""Based on this answer, extract the information into the requested structured format.
+        structured_prompt = f"""Your task is to extract specific information from the answer below and format it as JSON.

-Answer: {answer}
+ANSWER TO EXTRACT FROM:
+\"\"\"
+{answer}
+\"\"\"

-JSON Schema to follow:
+REQUIRED OUTPUT FORMAT - Extract the following fields from the answer above:
+{fields_text}
+
+JSON Schema:
 ```json
 {schema_str}
 ```

-Return ONLY a valid JSON object that matches this exact schema. Pay special attention to field types:
- "type": "array" means the value must be a JSON array/list, NOT a string
- "type": "string" means the value must be a string
- "type": "object" means the value must be a JSON object
+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

-Do not include any explanation, only the JSON object."""
+OUTPUT:"""

        structured_result, usage = await llm_config.call(
            messages=[
                {
                    "role": "system",
-                    "content": "Extract structured data from the given answer. Return only valid JSON matching the provided schema exactly.",
+                    "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},
            ],
@@ -188,6 +247,12 @@ Do not include any explanation, only the JSON object."""
            # 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

@@ -196,6 +261,46 @@ Do not include any explanation, only the JSON object."""
        return None, 0, 0


+_TIKTOKEN_ENCODING = tiktoken.get_encoding("cl100k_base")
+
+
+def _count_messages_tokens(messages: list[dict[str, Any]]) -> int:
+    """Estimate the token count of the messages list using cl100k_base encoding."""
+    total = 0
+    for msg in messages:
+        content = msg.get("content") or ""
+        if isinstance(content, str):
+            total += len(_TIKTOKEN_ENCODING.encode(content))
+        elif isinstance(content, list):
+            for part in content:
+                if isinstance(part, dict) and isinstance(part.get("text"), str):
+                    total += len(_TIKTOKEN_ENCODING.encode(part["text"]))
+        # Tool call arguments and results also count
+        for tc in msg.get("tool_calls") or []:
+            if isinstance(tc, dict):
+                func = tc.get("function", {})
+                total += len(_TIKTOKEN_ENCODING.encode(func.get("arguments", "")))
+    return total
+
+
+def _is_context_overflow_error(exc: Exception) -> bool:
+    """Return True if the exception signals the LLM context window was exceeded."""
+    msg = str(exc).lower()
+    return any(
+        phrase in msg
+        for phrase in (
+            "context_length_exceeded",
+            "context length exceeded",
+            "maximum context length",
+            "prompt_too_long",
+            "prompt is too long",
+            "resource_exhausted",
+            "input is too long",
+            "too many tokens",
+        )
+    )
+
+
 async def run_reflect_agent(
    llm_config: "LLMProvider",
    bank_id: str,
@@ -203,7 +308,7 @@ async def run_reflect_agent(
    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]]],
+    recall_fn: Callable[[str, int, 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,
@@ -212,6 +317,7 @@ async def run_reflect_agent(
    directives: list[dict[str, Any]] | None = None,
    has_mental_models: bool = False,
    budget: str | None = None,
+    max_context_tokens: int = 100_000,
 ) -> ReflectAgentResult:
    """
    Execute the reflect agent loop using native tool calling.
@@ -319,19 +425,22 @@ async def run_reflect_agent(
            f"total={elapsed_ms}ms"
        )

+    consecutive_errors = 0
    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)
+            prompt = build_final_prompt(
+                query, context_history, bank_profile, context, max_context_tokens=max_context_tokens
+            )
            llm_start = time.time()
            response, usage = await llm_config.call(
                messages=[
                    {"role": "system", "content": FINAL_SYSTEM_PROMPT},
                    {"role": "user", "content": prompt},
                ],
-                scope="reflect_agent_final",
+                scope="reflect",
                max_completion_tokens=max_tokens,
                return_usage=True,
            )
@@ -369,17 +478,91 @@ async def run_reflect_agent(
                directives_applied=directives_applied,
            )

+        # Proactive context-window guard: if accumulated messages would exceed the
+        # configured token budget, bail out early and synthesize from what we have.
+        estimated_tokens = _count_messages_tokens(messages)
+        if estimated_tokens >= max_context_tokens and (
+            bool(available_memory_ids) or bool(available_mental_model_ids) or bool(available_observation_ids)
+        ):
+            logger.warning(
+                f"[REFLECT {reflect_id}] Context budget exceeded on iteration {iteration + 1}: "
+                f"~{estimated_tokens} tokens >= {max_context_tokens} limit. Forcing final synthesis."
+            )
+            prompt = build_final_prompt(
+                query, context_history, bank_profile, context, max_context_tokens=max_context_tokens
+            )
+            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())
+
+            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()

+        # Determine tool_choice for this iteration.
+        # Force the full hierarchical retrieval path before allowing auto:
+        # With mental models:
+        #   0 → search_mental_models, 1 → search_observations, 2 → recall, 3+ → auto
+        # Without mental models:
+        #   0 → search_observations, 1 → recall, 2+ → auto
+        if iteration == 0 and has_mental_models:
+            iter_tool_choice: str | dict = {"type": "function", "function": {"name": "search_mental_models"}}
+        elif iteration == 0:
+            iter_tool_choice = {"type": "function", "function": {"name": "search_observations"}}
+        elif iteration == 1 and has_mental_models:
+            iter_tool_choice = {"type": "function", "function": {"name": "search_observations"}}
+        elif iteration == 1 or (iteration == 2 and has_mental_models):
+            iter_tool_choice = {"type": "function", "function": {"name": "recall"}}
+        else:
+            iter_tool_choice = "auto"
+
        try:
            result = await llm_config.call_with_tools(
                messages=messages,
                tools=tools,
-                scope="reflect_agent",
-                tool_choice="required" if iteration == 0 else "auto",  # Force tool use on first iteration
+                scope="reflect_tool_call",
+                tool_choice=iter_tool_choice,
            )
            llm_duration = int((time.time() - llm_start) * 1000)
+            consecutive_errors = 0
            total_input_tokens += result.input_tokens
            total_output_tokens += result.output_tokens
            llm_trace.append(
@@ -393,22 +576,32 @@ async def run_reflect_agent(

        except Exception as e:
            err_duration = int((time.time() - llm_start) * 1000)
+            consecutive_errors += 1
            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:
+            # Context overflow errors must never be retried — retrying would only make them worse.
+            # Skip straight to final synthesis with whatever evidence we have.
+            if _is_context_overflow_error(e):
+                logger.warning(
+                    f"[REFLECT {reflect_id}] Context window exceeded on iteration {iteration + 1}, "
+                    "forcing final synthesis from gathered evidence."
+                )
+            # For other errors: retry if no evidence yet (but cap consecutive errors to avoid long hangs)
+            elif not has_gathered_evidence and iteration < max_iterations - 1 and consecutive_errors < 2:
                continue
-            prompt = build_final_prompt(query, context_history, bank_profile, context)
+            prompt = build_final_prompt(
+                query, context_history, bank_profile, context, max_context_tokens=max_context_tokens
+            )
            llm_start = time.time()
            response, usage = await llm_config.call(
                messages=[
                    {"role": "system", "content": FINAL_SYSTEM_PROMPT},
                    {"role": "user", "content": prompt},
                ],
-                scope="reflect_agent_final",
+                scope="reflect",
                max_completion_tokens=max_tokens,
                return_usage=True,
            )
@@ -472,14 +665,16 @@ async def run_reflect_agent(
                    directives_applied=directives_applied,
                )
            # Empty response, force final
-            prompt = build_final_prompt(query, context_history, bank_profile, context)
+            prompt = build_final_prompt(
+                query, context_history, bank_profile, context, max_context_tokens=max_context_tokens
+            )
            llm_start = time.time()
            response, usage = await llm_config.call(
                messages=[
                    {"role": "system", "content": FINAL_SYSTEM_PROMPT},
                    {"role": "user", "content": prompt},
                ],
-                scope="reflect_agent_final",
+                scope="reflect",
                max_completion_tokens=max_tokens,
                return_usage=True,
            )
@@ -546,23 +741,30 @@ async def run_reflect_agent(
                )
                continue

-            # Process done tool
-            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,
-            )
+            # 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)]
@@ -722,14 +924,16 @@ async def _process_done_tool(
    """Process the done tool call and return the result."""
    args = done_call.arguments

-    answer = args.get("answer", "").strip()
+    # 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]
+    used_memory_ids = [mid for mid in (args.get("memory_ids") or []) if mid in available_memory_ids]
+    used_mental_model_ids = [mid for mid in (args.get("mental_model_ids") or []) if mid in available_mental_model_ids]
+    used_observation_ids = [oid for oid in (args.get("observation_ids") or []) if oid in available_observation_ids]

    # Generate structured output if schema provided
    structured_output = None
@@ -765,21 +969,71 @@ 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]]],
+    recall_fn: Callable[[str, int, 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."""
-    start = time.time()
-    result = await _execute_tool(
-        tc.name,
-        tc.arguments,
-        search_mental_models_fn,
-        search_observations_fn,
-        recall_fn,
-        expand_fn,
-    )
-    duration_ms = int((time.time() - start) * 1000)
-    return result, duration_ms
+    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(
@@ -787,7 +1041,7 @@ async def _execute_tool(
    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]]],
+    recall_fn: Callable[[str, int, int], Awaitable[dict[str, Any]]],
    expand_fn: Callable[[list[str], str], Awaitable[dict[str, Any]]],
 ) -> dict[str, Any]:
    """Execute a single tool by name."""
@@ -798,22 +1052,23 @@ async def _execute_tool(
        query = args.get("query")
        if not query:
            return {"error": "search_mental_models requires a query parameter"}
-        max_results = args.get("max_results") or 5
+        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(args.get("max_tokens") or 5000, 1000)  # Default 5000, min 1000
+        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(args.get("max_tokens") or 2048, 1000)  # Default 2048, min 1000
-        return await recall_fn(query, max_tokens)
+        max_tokens = max(int(args.get("max_tokens") or 2048), 1000)  # Default 2048, min 1000
+        max_chunk_tokens = max(int(args.get("max_chunk_tokens") or 1000), 1000)  # Always enabled, min 1000
+        return await recall_fn(query, max_tokens, max_chunk_tokens)

    elif tool_name == "expand":
        memory_ids = args.get("memory_ids", [])
@@ -831,19 +1086,19 @@ def _summarize_input(tool_name: str, args: dict[str, Any]) -> str:
    if tool_name == "search_mental_models":
        query = args.get("query", "")
        query_preview = f"'{query[:30]}...'" if len(query) > 30 else f"'{query}'"
-        max_results = args.get("max_results") or 5
+        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(args.get("max_tokens") or 5000, 1000)
+        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(args.get("max_tokens") or 2048, 1000)
-        return f"(query={query_preview}, max_tokens={max_tokens})"
+        max_tokens = max(int(args.get("max_tokens") or 2048), 1000)
+        max_chunk_tokens = max(int(args.get("max_chunk_tokens") or 1000), 1000)
+        return f"(query={query_preview}, max_tokens={max_tokens}, max_chunk_tokens={max_chunk_tokens})"
    elif tool_name == "expand":
        memory_ids = args.get("memory_ids", [])
        depth = args.get("depth", "chunk")
@@ -31,7 +31,7 @@ class ReflectAction(BaseModel):
        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="Plain text answer for done action (no markdown)")
+    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"
    )
@@ -10,59 +10,30 @@ The reflect agent uses hierarchical retrieval:
 import json
 from typing import Any

+import tiktoken
+
+_TIKTOKEN_ENCODING = tiktoken.get_encoding("cl100k_base")
+
+# Fraction of max_context_tokens reserved for tool results in the final synthesis prompt.
+# The remainder covers the system prompt, question, bank context, and output tokens.
+_FINAL_PROMPT_CONTEXT_FRACTION = 0.8
+

 def _extract_directive_rules(directives: list[dict[str, Any]]) -> list[str]:
-    """
-    Extract directive rules as a list of strings.
-
-    Args:
-        directives: List of directives with name and content
-
-    Returns:
-        List of directive rule strings
-    """
+    """Extract directive rules as a list of strings."""
    rules = []
    for directive in directives:
-        directive_name = directive.get("name", "")
-        # New format: directives have direct content field
+        name = directive.get("name", "")
        content = directive.get("content", "")
        if content:
-            if directive_name:
-                rules.append(f"**{directive_name}**: {content}")
-            else:
-                rules.append(content)
-        else:
-            # Legacy format: check for observations
-            observations = directive.get("observations", [])
-            if observations:
-                for obs in observations:
-                    # Support both Pydantic Observation objects and dicts
-                    if hasattr(obs, "title"):
-                        title = obs.title
-                        obs_content = obs.content
-                    else:
-                        title = obs.get("title", "")
-                        obs_content = obs.get("content", "")
-                    if title and obs_content:
-                        rules.append(f"**{title}**: {obs_content}")
-                    elif obs_content:
-                        rules.append(obs_content)
-            elif directive_name:
-                # Fallback to description
-                desc = directive.get("description", "")
-                if desc:
-                    rules.append(f"**{directive_name}**: {desc}")
+            rules.append(f"**{name}**: {content}" if name else content)
    return rules


 def build_directives_section(directives: list[dict[str, Any]]) -> str:
-    """
-    Build the directives section for the system prompt.
+    """Build the directives section for the system prompt.

    Directives are hard rules that MUST be followed in all responses.
-
-    Args:
-        directives: List of directive mental models with observations
    """
    if not directives:
        return ""
@@ -148,7 +119,15 @@ def build_system_prompt_for_tools(

    parts = []

-    # Inject directives at the VERY START for maximum prominence
+    # Anti-hallucination rule at the very top
+    parts.extend(
+        [
+            "CRITICAL: You MUST ONLY use information from retrieved tool results. NEVER make up names, people, events, or entities.",
+            "",
+        ]
+    )
+
+    # Inject directives after anti-hallucination rule
    if directives:
        parts.append(build_directives_section(directives))

@@ -161,8 +140,14 @@ def build_system_prompt_for_tools(

    parts.extend(
        [
+            "## LANGUAGE RULE (default - directives take precedence)",
+            "- By default, detect the language of the user's question and respond in that SAME language.",
+            "- If the question is in Chinese, respond in Chinese. If in Japanese, respond in Japanese.",
+            "- IMPORTANT: The DIRECTIVES section above has HIGHER PRIORITY than this rule.",
+            "  If a directive specifies a language (e.g. 'Always respond in French'), follow the directive.",
+            "",
            "## CRITICAL RULES",
-            "- You must NEVER fabricate information that has no basis in retrieved data",
+            "- ONLY use information from tool results - no external knowledge or guessing",
            "- You SHOULD synthesize, infer, and reason from the retrieved memories",
            "- You MUST search before saying you don't have information",
            "",
@@ -197,6 +182,7 @@ def build_system_prompt_for_tools(
                "### 3. RAW FACTS (recall) - Ground Truth",
                "- Individual memories (world facts and experiences)",
                "- Use when: no mental models/observations exist, they're stale, or you need specific details",
+                "- MANDATORY: If search_mental_models and search_observations both return 0 results, you MUST call recall() before giving up",
                "- This is the source of truth that other levels are built from",
                "",
            ]
@@ -214,6 +200,7 @@ def build_system_prompt_for_tools(
                "### 2. RAW FACTS (recall) - Ground Truth",
                "- Individual memories (world facts and experiences)",
                "- Use when: no observations exist, they're stale, or you need specific details",
+                "- MANDATORY: If search_observations returns 0 results or count=0, you MUST call recall() before giving up",
                "- This is the source of truth that observations are built from",
                "",
            ]
@@ -291,7 +278,7 @@ def build_system_prompt_for_tools(
        parts.extend(
            [
                "1. First, try search_observations() - check for consolidated knowledge",
-                "2. If observations are stale OR you need specific details, use recall() for raw facts",
+                "2. If search_observations returns 0 results OR observations are stale, you MUST call recall() for raw facts",
                "3. Use expand() if you need more context on specific memories",
                "4. When ready, call done() with your answer and supporting IDs",
            ]
@@ -300,11 +287,14 @@ def build_system_prompt_for_tools(
    parts.extend(
        [
            "",
-            "## Output Format: Plain Text Answer",
-            "Call done() with a plain text 'answer' field.",
-            "- Do NOT use markdown formatting",
+            "## Output Format: Well-Formatted Markdown Answer",
+            "Call done() with a well-formatted markdown 'answer' field.",
+            "- USE markdown formatting for structure (headers, lists, bold, italic, code blocks, tables, etc.)",
+            "- CRITICAL: Add blank lines before and after block elements (tables, code blocks, lists)",
+            "- Format for clarity and readability with proper spacing and hierarchy",
            "- NEVER include memory IDs, UUIDs, or 'Memory references' in the answer text",
            "- Put IDs ONLY in the memory_ids/mental_model_ids/observation_ids arrays, not in the answer",
+            "- CRITICAL: This is a NON-CONVERSATIONAL system. NEVER ask follow-up questions, offer further assistance, or suggest next steps. Your answer must be complete and self-contained. The user cannot reply.",
        ]
    )

@@ -412,6 +402,7 @@ def build_final_prompt(
    context_history: list[dict],
    bank_profile: dict,
    additional_context: str | None = None,
+    max_context_tokens: int = 100_000,
 ) -> str:
    """Build the final prompt when forcing a text response (no tools)."""
    parts = []
@@ -441,18 +432,32 @@ def build_final_prompt(
    if additional_context:
        parts.append(f"\n## Additional Context\n{additional_context}")

-    # Tool call history
+    # Tool call history — include as many entries as fit within the token budget,
+    # preferring the most recent calls (they tend to be the most targeted).
    if context_history:
        parts.append("\n## Retrieved Data (synthesize and reason from this data)")
-        for entry in context_history:
+        token_budget = int(max_context_tokens * _FINAL_PROMPT_CONTEXT_FRACTION)
+        # Render entries newest-first, then reverse so the prompt reads chronologically.
+        rendered: list[str] = []
+        truncated = False
+        for entry in reversed(context_history):
            tool = entry["tool"]
            output = entry["output"]
-            # Format as proper JSON for LLM readability
            try:
                output_str = json.dumps(output, indent=2, default=str)
            except (TypeError, ValueError):
                output_str = str(output)
-            parts.append(f"\n### From {tool}:\n```json\n{output_str}\n```")
+            block = f"\n### From {tool}:\n```json\n{output_str}\n```"
+            block_tokens = len(_TIKTOKEN_ENCODING.encode(block))
+            if block_tokens > token_budget:
+                truncated = True
+                break
+            rendered.append(block)
+            token_budget -= block_tokens
+        for block in reversed(rendered):
+            parts.append(block)
+        if truncated:
+            parts.append("\n*Note: Some earlier tool results were omitted to stay within the context window.*")
    else:
        parts.append("\n## Retrieved Data\nNo data was retrieved.")

@@ -463,21 +468,43 @@ def build_final_prompt(
    parts.append(
        "\n## Instructions\n"
        "Provide a thoughtful answer by synthesizing and reasoning from the retrieved data above. "
-        "You can make reasonable inferences from the memories, but don't completely fabricate information."
+        "You can make reasonable inferences from the memories, but don't completely fabricate information. "
        "If the exact answer isn't stated, use what IS stated to give the best possible answer. "
-        "Only say 'I don't have information' if the retrieved data is truly unrelated to the question."
+        "Only say 'I don't have information' if the retrieved data is truly unrelated to the question.\n\n"
+        "IMPORTANT: Output ONLY the final answer. Do NOT include meta-commentary like "
+        '"I\'ll search..." or "Let me analyze...". Do NOT explain your reasoning process. '
+        "Just provide the direct synthesized answer."
    )

    return "\n".join(parts)


-FINAL_SYSTEM_PROMPT = """You are a thoughtful assistant that synthesizes answers from retrieved memories.
+FINAL_SYSTEM_PROMPT = """CRITICAL: You MUST ONLY use information from retrieved tool results. NEVER make up names, people, events, or entities.
+
+You are a thoughtful assistant that synthesizes answers from retrieved memories.

 Your approach:
 - Reason over the retrieved memories to answer the question
 - Make reasonable inferences when the exact answer isn't explicitly stated
 - Connect related memories to form a complete picture
 - Be helpful - if you have related information, use it to give the best possible answer
+- ONLY use information from tool results - no external knowledge or guessing

 Only say "I don't have information" if the retrieved data is truly unrelated to the question.
-Do NOT fabricate information that has no basis in the retrieved data."""
+
+FORMATTING: Use proper markdown formatting in your answer:
+- Headers (##, ###) for sections
+- Lists (bullet or numbered) for enumerations
+- Bold/italic for emphasis
+- Tables with proper syntax (ensure blank line before and after)
+- Code blocks where appropriate
+- CRITICAL: Always add blank lines before and after block elements (tables, code blocks, lists)
+- Proper spacing between sections
+
+CRITICAL: Output ONLY the final synthesized answer. Do NOT include:
+- Meta-commentary about what you're doing ("I'll search...", "Let me analyze...")
+- Explanations of your reasoning process
+- Descriptions of your approach
+Just provide the direct answer with proper markdown formatting.
+
+CRITICAL: This is a NON-CONVERSATIONAL system. NEVER ask follow-up questions, offer to search again, suggest alternatives, or end with anything like "Would you like me to..." or "Let me know if...". The user cannot reply. Your answer must be complete and self-contained."""
@@ -9,7 +9,7 @@ Implements hierarchical retrieval:

 import logging
 import uuid
-from datetime import datetime, timedelta, timezone
+from datetime import datetime, timezone
 from typing import TYPE_CHECKING, Any

 if TYPE_CHECKING:
@@ -20,9 +20,6 @@ if TYPE_CHECKING:

 logger = logging.getLogger(__name__)

-# Observation is considered stale if not updated in this many days
-STALE_THRESHOLD_DAYS = 7
-

 async def tool_search_mental_models(
    conn: "Connection",
@@ -33,6 +30,7 @@ async def tool_search_mental_models(
    tags: list[str] | None = None,
    tags_match: str = "any",
    exclude_ids: list[str] | None = None,
+    pending_consolidation: int = 0,
 ) -> dict[str, Any]:
    """
    Search user-curated mental models by semantic similarity.
@@ -54,22 +52,21 @@ async def tool_search_mental_models(
        Dict with matching mental models including content and freshness info
    """
    from ..memory_engine import fq_table
+    from ..search.tags import build_tags_where_clause

    # Build filters dynamically
    filters = ""
    params: list[Any] = [bank_id, str(query_embedding), max_results]
    next_param = 4

+    # Use the centralized tag filtering logic
    if tags:
-        if tags_match == "all":
-            filters += f" AND tags @> ${next_param}::varchar[]"
-        else:
-            filters += f" AND (tags && ${next_param}::varchar[] OR tags IS NULL OR tags = '{{}}')"
-        params.append(tags)
-        next_param += 1
+        tag_clause, tag_params, next_param = build_tags_where_clause(tags, param_offset=next_param, match=tags_match)
+        filters += f" {tag_clause}"
+        params.extend(tag_params)

    if exclude_ids:
-        filters += f" AND id != ALL(${next_param}::uuid[])"
+        filters += f" AND id != ALL(${next_param}::text[])"
        params.append(exclude_ids)
        next_param += 1

@@ -88,7 +85,6 @@ async def tool_search_mental_models(
        *params,
    )

-    now = datetime.now(timezone.utc)
    mental_models = []

    for row in rows:
@@ -96,11 +92,10 @@ async def tool_search_mental_models(
        if last_refreshed_at and last_refreshed_at.tzinfo is None:
            last_refreshed_at = last_refreshed_at.replace(tzinfo=timezone.utc)

-        # Calculate freshness
-        is_stale = False
-        if last_refreshed_at:
-            age = now - last_refreshed_at
-            is_stale = age > timedelta(days=STALE_THRESHOLD_DAYS)
+        # A mental model is stale when there are memories that haven't been consolidated yet —
+        # the same signal used for observations staleness.
+        is_stale = pending_consolidation > 0
+        staleness_reason = f"{pending_consolidation} memories pending consolidation" if is_stale else None

        mental_models.append(
            {
@@ -111,6 +106,7 @@ async def tool_search_mental_models(
                "relevance": round(row["relevance"], 4),
                "updated_at": last_refreshed_at.isoformat() if last_refreshed_at else None,
                "is_stale": is_stale,
+                "staleness_reason": staleness_reason,
            }
        )

@@ -133,7 +129,7 @@ async def tool_search_observations(
    pending_consolidation: int = 0,
 ) -> dict[str, Any]:
    """
-    Search consolidated observations using recall with include_observations.
+    Search consolidated observations using recall with include_source_facts.

    Observations are auto-generated from memories. Returns freshness info
    so the agent knows if it should also verify with recall().
@@ -150,72 +146,24 @@ async def tool_search_observations(
        pending_consolidation: Number of memories waiting to be consolidated

    Returns:
-        Dict with matching observations including freshness info
+        Dict with matching observations including freshness info and source memories
    """
-    from ..memory_engine import fq_table
-
-    # Use recall to search observations (they come back in results field when fact_type=["observation"])
    result = await memory_engine.recall_async(
        bank_id=bank_id,
        query=query,
-        fact_type=["observation"],  # Only retrieve observations
-        max_tokens=max_tokens,  # Token budget controls how many observations are returned
+        fact_type=["observation"],
+        max_tokens=max_tokens,
        enable_trace=False,
        request_context=request_context,
        tags=tags,
        tags_match=tags_match,
+        include_source_facts=True,
+        max_source_facts_tokens=-1,  # No token limit — include all source facts
        _connection_budget=1,
        _quiet=True,
    )

-    observations = []
-
-    # When fact_type=["observation"], results come back in `results` field as MemoryFact objects
-    # We need to fetch additional fields (proof_count, source_memory_ids) from the database
-    if result.results:
-        obs_ids = [m.id for m in result.results]
-
-        # Fetch proof_count and source_memory_ids for these observations
-        pool = await memory_engine._get_pool()
-        async with pool.acquire() as conn:
-            obs_rows = await conn.fetch(
-                f"""
-                SELECT id, proof_count, source_memory_ids
-                FROM {fq_table("memory_units")}
-                WHERE id = ANY($1::uuid[])
-                """,
-                obs_ids,
-            )
-            obs_data = {str(row["id"]): row for row in obs_rows}
-
-        for m in result.results:
-            # Get additional data from DB lookup
-            extra = obs_data.get(m.id, {})
-            proof_count = extra.get("proof_count", 1) if extra else 1
-            source_ids = extra.get("source_memory_ids", []) if extra else []
-            # Convert UUIDs to strings
-            source_memory_ids = [str(sid) for sid in (source_ids or [])]
-
-            # Determine staleness
-            is_stale = False
-            staleness_reason = None
-            if pending_consolidation > 0:
-                is_stale = True
-                staleness_reason = f"{pending_consolidation} memories pending consolidation"
-
-            observations.append(
-                {
-                    "id": str(m.id),
-                    "text": m.text,
-                    "proof_count": proof_count,
-                    "source_memory_ids": source_memory_ids,
-                    "tags": m.tags or [],
-                    "is_stale": is_stale,
-                    "staleness_reason": staleness_reason,
-                }
-            )
-
-    # Return freshness info (more understandable than raw pending_consolidation count)
+    is_stale = pending_consolidation > 0
    if pending_consolidation == 0:
        freshness = "up_to_date"
    elif pending_consolidation < 10:
@@ -225,8 +173,10 @@ async def tool_search_observations(

    return {
        "query": query,
-        "count": len(observations),
-        "observations": observations,
+        "count": len(result.results),
+        "observations": [m.model_dump() for m in result.results],
+        "source_facts": {k: v.model_dump() for k, v in (result.source_facts or {}).items()},
+        "is_stale": is_stale,
        "freshness": freshness,
    }

@@ -237,10 +187,10 @@ async def tool_recall(
    query: str,
    request_context: "RequestContext",
    max_tokens: int = 2048,
-    max_results: int = 50,
    tags: list[str] | None = None,
    tags_match: str = "any",
    connection_budget: int = 1,
+    max_chunk_tokens: int = 1000,
 ) -> dict[str, Any]:
    """
    Search memories using TEMPR retrieval.
@@ -254,18 +204,19 @@ async def tool_recall(
        query: Search query
        request_context: Request context for authentication
        max_tokens: Maximum tokens for results (default 2048)
-        max_results: Maximum number of results
        tags: Filter by tags (includes untagged memories)
        tags_match: How to match tags - "any" (OR), "all" (AND), or "exact"
        connection_budget: Max DB connections for this recall (default 1 for internal ops)
+        max_chunk_tokens: Maximum tokens for raw source chunk text (default 1000, always included)

    Returns:
-        Dict with list of matching memories
+        Dict with list of matching memories including raw chunk text
    """
+    include_chunks = True
    result = await memory_engine.recall_async(
        bank_id=bank_id,
        query=query,
-        fact_type=["experience", "world"],  # Exclude opinions and observations
+        fact_type=["experience", "world"],
        max_tokens=max_tokens,
        enable_trace=False,
        request_context=request_context,
@@ -273,24 +224,14 @@ async def tool_recall(
        tags_match=tags_match,
        _connection_budget=connection_budget,
        _quiet=True,  # Suppress logging for internal operations
+        include_chunks=include_chunks,
+        max_chunk_tokens=max_chunk_tokens,
    )

-    memories = []
-    for m in result.results[:max_results]:
-        memories.append(
-            {
-                "id": str(m.id),
-                "text": m.text,
-                "type": m.fact_type,
-                "entities": m.entities or [],
-                "occurred": m.occurred_start,  # Already ISO format string
-            }
-        )
-
    return {
        "query": query,
-        "count": len(memories),
-        "memories": memories,
+        "memories": [m.model_dump() for m in result.results],
+        "chunks": {k: v.model_dump() for k, v in (result.chunks or {}).items()},
    }


@@ -47,7 +47,8 @@ TOOL_SEARCH_OBSERVATIONS = {
        "description": (
            "Search consolidated observations (auto-generated knowledge). These are automatically "
            "synthesized from memories. Returns observations with freshness info (updated_at, is_stale). "
-            "If an observation is STALE, you should ALSO use recall() to verify with current facts."
+            "If an observation is STALE, you should ALSO use recall() to verify with current facts. "
+            "IMPORTANT: If search_mental_models is available, you MUST call it FIRST before using this tool."
        ),
        "parameters": {
            "type": "object",
@@ -95,6 +96,10 @@ TOOL_RECALL = {
                    "type": "integer",
                    "description": "Optional limit on result size (default 2048). Use higher values for broader searches.",
                },
+                "max_chunk_tokens": {
+                    "type": "integer",
+                    "description": "Maximum tokens for raw source chunk text included alongside each memory fact (default 1000, min 1000). Chunks provide the surrounding context the fact was extracted from. Increase for broader context.",
+                },
            },
            "required": ["reason", "query"],
        },
@@ -139,7 +144,7 @@ TOOL_DONE_ANSWER = {
            "properties": {
                "answer": {
                    "type": "string",
-                    "description": "Your response as plain text. Do NOT use markdown formatting. NEVER include memory IDs, UUIDs, or 'Memory references' in this text - put IDs only in memory_ids array.",
+                    "description": "Your response as well-formatted markdown. Use headers, lists, bold/italic, and code blocks for clarity. NEVER include memory IDs, UUIDs, or 'Memory references' in this text - put IDs only in memory_ids array. LANGUAGE: By default, write in the SAME language as the user's question. However, if a language directive in the system prompt specifies a different language, follow that directive instead.",
                },
                "memory_ids": {
                    "type": "array",
@@ -190,7 +195,11 @@ def _build_done_tool_with_directives(directive_rules: list[str]) -> dict:
                "properties": {
                    "answer": {
                        "type": "string",
-                        "description": "Your response as plain text. Do NOT use markdown formatting. NEVER include memory IDs, UUIDs, or 'Memory references' in this text - put IDs only in memory_ids array.",
+                        "description": (
+                            "Your response as well-formatted markdown. Use headers, lists, bold/italic, and code blocks for clarity. "
+                            "NEVER include memory IDs, UUIDs, or 'Memory references' in this text - put IDs only in memory_ids array. "
+                            f"MANDATORY: Your answer MUST comply with ALL directives:\n{rules_list}"
+                        ),
                    },
                    "memory_ids": {
                        "type": "array",
@@ -159,6 +159,10 @@ class MemoryFact(BaseModel):
        None, description="ID of the chunk this fact was extracted from (format: bank_id_document_id_chunk_index)"
    )
    tags: list[str] | None = Field(None, description="Visibility scope tags associated with this fact")
+    source_fact_ids: list[str] | None = Field(
+        None,
+        description="IDs of source facts this observation was derived from (observation type only, when source_facts is enabled)",
+    )


 class ChunkInfo(BaseModel):
@@ -226,6 +230,9 @@ class RecallResult(BaseModel):
    chunks: dict[str, ChunkInfo] | None = Field(
        None, description="Chunks for facts, keyed by '{document_id}_{chunk_index}'"
    )
+    source_facts: dict[str, MemoryFact] | None = Field(
+        None, description="Source facts for observation-type results, keyed by fact ID"
+    )


 class ReflectResult(BaseModel):
@@ -263,7 +270,6 @@ class ReflectResult(BaseModel):
                        }
                    ],
                },
-                "new_opinions": ["Machine learning has great potential in healthcare"],
                "structured_output": {"summary": "ML in healthcare", "confidence": 0.9},
                "usage": {"input_tokens": 1500, "output_tokens": 500, "total_tokens": 2000},
            }
@@ -272,9 +278,8 @@ class ReflectResult(BaseModel):

    text: str = Field(description="The formulated answer text")
    based_on: dict[str, Any] = Field(
-        description="Facts used to formulate the answer, organized by type (world, experience, opinion, mental_models, directives)"
+        description="Facts used to formulate the answer, organized by type (world, experience, mental_models, directives)"
    )
-    new_opinions: list[str] = Field(default_factory=list, description="List of newly formed opinions during reflection")
    structured_output: dict[str, Any] | None = Field(
        default=None,
        description="Structured output parsed according to the provided response schema. Only present when response_schema was provided.",
@@ -297,24 +302,6 @@ class ReflectResult(BaseModel):
    )


-class Opinion(BaseModel):
-    """
-    An opinion with confidence score.
-
-    Opinions represent the bank's formed perspectives on topics,
-    with a confidence level indicating strength of belief.
-    """
-
-    model_config = ConfigDict(
-        json_schema_extra={
-            "example": {"text": "Machine learning has great potential in healthcare", "confidence": 0.85}
-        }
-    )
-
-    text: str = Field(description="The opinion text")
-    confidence: float = Field(description="Confidence score between 0.0 and 1.0")
-
-
 class EntityObservation(BaseModel):
    """
    An observation about an entity.
@@ -5,7 +5,6 @@ This package contains modular components for the retain operation:
 - types: Type definitions for retain pipeline
 - fact_extraction: Extract facts from content
 - embedding_processing: Augment texts and generate embeddings
- deduplication: Check for duplicate facts
 - entity_processing: Process and resolve entities
 - link_creation: Create temporal, semantic, entity, and causal links
 - chunk_storage: Handle chunk storage
@@ -14,7 +13,6 @@ This package contains modular components for the retain operation:

 from . import (
    chunk_storage,
-    deduplication,
    embedding_processing,
    entity_processing,
    fact_extraction,
@@ -35,7 +33,6 @@ __all__ = [
    # Modules
    "fact_extraction",
    "embedding_processing",
-    "deduplication",
    "entity_processing",
    "link_creation",
    "chunk_storage",
@@ -1,85 +0,0 @@
-"""
-Deduplication logic for retain pipeline.
-
-Checks for duplicate facts using semantic similarity and temporal proximity.
-"""
-
-import logging
-from collections import defaultdict
-from datetime import UTC
-
-from .types import ProcessedFact
-
-logger = logging.getLogger(__name__)
-
-
-async def check_duplicates_batch(conn, bank_id: str, facts: list[ProcessedFact], duplicate_checker_fn) -> list[bool]:
-    """
-    Check which facts are duplicates using batched time-window queries.
-
-    Groups facts by 12-hour time buckets to efficiently check for duplicates
-    within a 24-hour window.
-
-    Args:
-        conn: Database connection
-        bank_id: Bank identifier
-        facts: List of ProcessedFact objects to check
-        duplicate_checker_fn: Async function(conn, bank_id, texts, embeddings, date, time_window_hours)
-                              that returns List[bool] indicating duplicates
-
-    Returns:
-        List of boolean flags (same length as facts) indicating if each fact is a duplicate
-    """
-    if not facts:
-        return []
-
-    # Group facts by event_date (rounded to 12-hour buckets) for efficient batching
-    time_buckets = defaultdict(list)
-    for idx, fact in enumerate(facts):
-        # Use occurred_start if available, otherwise use mentioned_at
-        # For deduplication purposes, we need a time reference
-        fact_date = fact.occurred_start if fact.occurred_start is not None else fact.mentioned_at
-
-        # Defensive: if both are None (shouldn't happen), use now()
-        if fact_date is None:
-            from datetime import datetime
-
-            fact_date = datetime.now(UTC)
-
-        # Round to 12-hour bucket to group similar times
-        bucket_key = fact_date.replace(hour=(fact_date.hour // 12) * 12, minute=0, second=0, microsecond=0)
-        time_buckets[bucket_key].append((idx, fact))
-
-    # Process each bucket in batch
-    all_is_duplicate = [False] * len(facts)
-
-    for bucket_date, bucket_items in time_buckets.items():
-        indices = [item[0] for item in bucket_items]
-        texts = [item[1].fact_text for item in bucket_items]
-        embeddings = [item[1].embedding for item in bucket_items]
-
-        # Check duplicates for this time bucket
-        dup_flags = await duplicate_checker_fn(conn, bank_id, texts, embeddings, bucket_date, time_window_hours=24)
-
-        # Map results back to original indices
-        for idx, is_dup in zip(indices, dup_flags):
-            all_is_duplicate[idx] = is_dup
-
-    return all_is_duplicate
-
-
-def filter_duplicates(facts: list[ProcessedFact], is_duplicate_flags: list[bool]) -> list[ProcessedFact]:
-    """
-    Filter out duplicate facts based on duplicate flags.
-
-    Args:
-        facts: List of ProcessedFact objects
-        is_duplicate_flags: Boolean flags indicating which facts are duplicates
-
-    Returns:
-        List of non-duplicate facts
-    """
-    if len(facts) != len(is_duplicate_flags):
-        raise ValueError(f"Mismatch between facts ({len(facts)}) and flags ({len(is_duplicate_flags)})")
-
-    return [fact for fact, is_dup in zip(facts, is_duplicate_flags) if not is_dup]
@@ -27,11 +27,21 @@ def augment_texts_with_dates(facts: list[ExtractedFact], format_date_fn) -> list
    """
    augmented_texts = []
    for fact in facts:
-        # Use occurred_start as the representative date
+        # Use occurred_start as the representative date, fall back to mentioned_at
        fact_date = fact.occurred_start or fact.mentioned_at
-        readable_date = format_date_fn(fact_date)
-        # Augment text with date for embedding (but store original text in DB)
-        augmented_text = f"{fact.fact_text} (happened in {readable_date})"
+        # Augment text with date and entity names for embedding (but store original text in DB)
+        # Entity names (including key:value labels) improve retrieval without polluting stored content
+        if fact_date is not None:
+            readable_date = format_date_fn(fact_date)
+            if fact.occurred_end and fact.occurred_end != fact.occurred_start:
+                readable_end = format_date_fn(fact.occurred_end)
+                augmented_text = f"{fact.fact_text} (happened from {readable_date} to {readable_end})"
+            else:
+                augmented_text = f"{fact.fact_text} (happened in {readable_date})"
+        else:
+            augmented_text = fact.fact_text
+        if fact.entities:
+            augmented_text = f"{augmented_text} [{', '.join(fact.entities)}]"
        augmented_texts.append(augmented_text)
    return augmented_texts

@@ -41,10 +41,9 @@ async def generate_embeddings_batch(embeddings_backend, texts: list[str]) -> lis
        List of embeddings in same order as input texts
    """
    try:
-        # Run embeddings in thread pool to avoid blocking event loop
        loop = asyncio.get_event_loop()
        embeddings = await loop.run_in_executor(
-            None,  # Use default thread pool
+            None,
            embeddings_backend.encode,
            texts,
        )
@@ -0,0 +1,194 @@
+"""
+Entity labels models and helpers for retain pipeline.
+
+Defines a controlled vocabulary of key:value classification labels
+(e.g., 'pedagogy:scaffolding', 'interest:active') that are extracted
+at retain time and stored as entities.
+"""
+
+from typing import Literal
+
+from pydantic import BaseModel, Field, create_model
+
+
+class LabelValue(BaseModel):
+    """A single allowed value for a label group."""
+
+    value: str
+    description: str = ""
+
+
+class LabelGroup(BaseModel):
+    """A label group (dimension) with its type and allowed values."""
+
+    key: str
+    description: str = ""
+    type: Literal["value", "multi-values", "text"] = "value"
+    optional: bool = True
+    tag: bool = False
+    values: list[LabelValue] = []
+
+
+class EntityLabelsConfig(BaseModel):
+    """Entity labels configuration for a bank (controlled vocabulary)."""
+
+    attributes: list[LabelGroup] = []
+
+
+def parse_entity_labels(raw: dict | list | None) -> EntityLabelsConfig | None:
+    """
+    Parse raw entity labels config into EntityLabelsConfig.
+
+    Accepts:
+    - None → returns None
+    - list → list of attribute dicts (each may use legacy free_values/multi_value or new type field)
+    - dict → {attributes: [...]}
+
+    Legacy migration (backward-compat):
+    - free_values=True          → type="text"
+    - multi_value=True          → type="multi-values"
+    - neither / free_values=False → type="value"
+
+    Args:
+        raw: Raw entity labels config from bank config
+
+    Returns:
+        EntityLabelsConfig or None if raw is None/empty
+    """
+    if raw is None:
+        return None
+
+    if isinstance(raw, list):
+        if not raw:
+            return None
+        attributes = [LabelGroup.model_validate(_migrate_label_group(a)) for a in raw]
+        return EntityLabelsConfig(attributes=attributes)
+
+    if isinstance(raw, dict):
+        attrs_raw = raw.get("attributes", [])
+        if not attrs_raw:
+            return None
+        attributes = [LabelGroup.model_validate(_migrate_label_group(a)) for a in attrs_raw]
+        return EntityLabelsConfig(attributes=attributes)
+
+    return None
+
+
+def _migrate_label_group(raw: dict) -> dict:
+    """Migrate legacy free_values/multi_value fields to the new type field."""
+    if not isinstance(raw, dict) or "type" in raw:
+        return raw
+    patched = dict(raw)
+    if patched.get("free_values"):
+        patched["type"] = "text"
+    elif patched.get("multi_value"):
+        patched["type"] = "multi-values"
+    else:
+        patched["type"] = "value"
+    # Remove legacy keys so Pydantic doesn't error on unknown fields
+    patched.pop("free_values", None)
+    patched.pop("multi_value", None)
+    return patched
+
+
+def build_labels_model(labels_cfg: EntityLabelsConfig) -> type[BaseModel] | None:
+    """
+    Build a dynamic Pydantic model for structured label extraction.
+
+    Each LabelGroup becomes a typed field based on its type:
+    - type="text"                        → str | None  (always optional)
+    - type="value",   optional=True      → Literal["v1","v2"] | None
+    - type="value",   optional=False     → Literal["v1","v2"]  (required)
+    - type="multi-values"                → list[Literal["v1","v2"]]
+
+    Args:
+        labels_cfg: Parsed EntityLabelsConfig
+
+    Returns:
+        Dynamic Pydantic model class, or None if no groups defined
+    """
+    fields: dict = {}
+    for group in labels_cfg.attributes:
+        if not group.key:
+            continue
+        description = group.description or group.key
+
+        if group.type == "text":
+            # Free-form: any string value accepted, always optional
+            fields[group.key] = (str | None, Field(default=None, description=description))
+        else:
+            # Enum-constrained: must have defined values
+            if not group.values:
+                continue
+            values = tuple(v.value for v in group.values if v.value)
+            if not values:
+                continue
+            # Literal[("v1", "v2")] is equivalent to Literal["v1", "v2"] in Python 3.11+
+            literal_type = Literal[values]  # type: ignore[valid-type]
+            if group.type == "multi-values":
+                fields[group.key] = (
+                    list[literal_type],  # type: ignore[valid-type]
+                    Field(default_factory=list, description=description),
+                )
+            elif group.optional:
+                fields[group.key] = (
+                    literal_type | None,  # type: ignore[valid-type]
+                    Field(default=None, description=description),
+                )
+            else:
+                fields[group.key] = (
+                    literal_type,  # type: ignore[valid-type]
+                    Field(description=description),
+                )
+
+    if not fields:
+        return None
+
+    return create_model("Labels", **fields)
+
+
+def is_label_entity(text: str, labels_cfg: EntityLabelsConfig, labels_lookup: set[str]) -> bool:
+    """
+    Return True if entity text belongs to any configured label group.
+
+    For enum groups: checks the pre-built lookup set.
+    For text groups: checks that the text starts with a known key prefix.
+    """
+    if text.lower() in labels_lookup:
+        return True
+    for group in labels_cfg.attributes:
+        if group.type == "text" and group.key and text.lower().startswith(f"{group.key.lower()}:"):
+            return True
+    return False
+
+
+def build_labels_lookup(labels_cfg: EntityLabelsConfig | list | None) -> set[str]:
+    """
+    Build a set of valid 'key:value' label strings (lowercase) for fast lookup.
+
+    Accepts either EntityLabelsConfig or raw list/None for backwards compatibility.
+
+    Args:
+        labels_cfg: EntityLabelsConfig, raw list of attribute dicts, or None
+
+    Returns:
+        Set of lowercase 'key:value' strings
+    """
+    if labels_cfg is None:
+        return set()
+
+    # Accept raw list/dict for backwards compatibility
+    if not isinstance(labels_cfg, EntityLabelsConfig):
+        parsed = parse_entity_labels(labels_cfg)
+        if parsed is None:
+            return set()
+        labels_cfg = parsed
+
+    valid = set()
+    for group in labels_cfg.attributes:
+        if group.type == "text":
+            continue  # No fixed vocabulary — all values accepted in post-processing
+        for v in group.values:
+            if group.key and v.value:
+                valid.add(f"{group.key}:{v.value}".lower())
+    return valid
@@ -20,6 +20,7 @@ async def process_entities_batch(
    facts: list[ProcessedFact],
    log_buffer: list[str] = None,
    user_entities_per_content: dict[int, list[dict]] = None,
+    entity_labels: list | None = None,
 ) -> list[EntityLink]:
    """
    Process entities for all facts and create entity links.
@@ -90,6 +91,7 @@ async def process_entities_batch(
        fact_dates,
        entities_per_fact,
        log_buffer,  # Pass log_buffer for detailed logging
+        entity_labels=entity_labels,
    )

    return entity_links
--- a/Show More
+++ b/Show More