Readwise Vector DB – Self-host your reading highlights search

Turn your Readwise library into a blazing-fast semantic search engine – complete with nightly syncs, vector search API, Prometheus metrics, and a streaming MCP server for LLM clients.

Table of Contents

• Readwise Vector DB – Self-host your reading highlights search

  • Table of Contents

  • Quick Start

  • Using Supabase Cloud

  • Deploy to Vercel in 3 Commands

    • Why SSE for MCP in Serverless?

  • Detailed Setup

    • Prerequisites

    • Environment Variables

    • Database & Migrations

    • Sync Commands (CLI)

  • Usage Examples

    • Vector Search (HTTP API)

    • Streaming Search (HTTP SSE)

    • Streaming Search (MCP TCP)

  • Architecture Overview

    • Docker + Local PostgreSQL (Default)

    • Vercel + Supabase (Cloud)

  • Development & Contribution

  • Maintainer Notes

  • License & Credits

Related MCP server: Readwise MCP

Quick Start

# ❶ Clone & install
git clone https://github.com/leonardsellem/readwise-vector-db.git
cd readwise-vector-db
poetry install --sync

# ❷ Boot DB & run the API (localhost:8000)
docker compose up -d db
poetry run uvicorn readwise_vector_db.api:app --reload

# ❸ Verify
curl http://127.0.0.1:8000/health     # → {"status":"ok"}
open http://127.0.0.1:8000/docs       # interactive swagger UI

Tip: Codespaces user? Click "Run → Open in Browser" after step ❷.

Using Supabase Cloud

Skip the local Docker setup and use a managed PostgreSQL with pgvector support:

# ❶ Create Supabase project at https://supabase.com/dashboard
# ❷ Enable pgvector extension in SQL Editor:
#   CREATE EXTENSION IF NOT EXISTS vector;

# ❸ Set up environment
export DB_BACKEND=supabase
export SUPABASE_DB_URL="postgresql://postgres:[password]@db.[project].supabase.co:6543/postgres?options=project%3D[project]"
export READWISE_TOKEN=xxxx
export OPENAI_API_KEY=sk-...

# ❹ Run migrations and start the API
poetry run alembic upgrade head
poetry run uvicorn readwise_vector_db.api:app --reload

# ❺ Initial sync
poetry run rwv sync --backfill

⚠️ Fail-fast behavior: The application will raise ValueError immediately on startup if SUPABASE_DB_URL is missing when DB_BACKEND=supabase.

Environment Variables Required:

• DB_BACKEND=supabase – Switches from local Docker to Supabase

• SUPABASE_DB_URL – Full PostgreSQL connection string from Supabase dashboard

• Standard variables: READWISE_TOKEN, OPENAI_API_KEY

Benefits:

• ✅ No Docker setup required

• ✅ Managed backups and scaling

• ✅ Built-in pgvector support

• ✅ Global edge network

• ✅ SSE streaming optimized – Connection pooling and sub-100ms query latency

Deploy to Vercel in 3 Commands

Deploy the FastAPI app as a serverless function with Supabase backend:

# ❶ Set up Vercel project
npm install -g vercel
vercel login
vercel link  # or vercel --confirm for new project

# ❶ Configure environment variables in Vercel dashboard or CLI:
vercel env add SUPABASE_DB_URL
vercel env add READWISE_TOKEN
vercel env add OPENAI_API_KEY

# ❸ Deploy
vercel --prod

Automatic Configuration:

• DEPLOY_TARGET=vercel – Automatically set by Vercel environment

• DB_BACKEND=supabase – Pre-configured in vercel.json

• Build process uses optimized vercel_build.sh script

Resource Limits:

• ⏱️ Build timeout: 90 seconds

• 💾 Memory limit: 1024MB during build

• 🚀 Function timeout: 30 seconds per request

SSE Streaming Support:

• ✅ HTTP-based MCP Server – /mcp/stream endpoint works seamlessly

• ✅ Real-time search results – Server-Sent Events for streaming responses

• ✅ Cold-start optimized – Sub-1s initialization, auto-scaling connections

• ✅ HTTP/2 multiplexing – Unlimited concurrent connections per client

GitHub Integration:

• Tagged releases (v*.*.*) automatically deploy to production

• Pull requests create preview deployments

• CI validates both Docker and Vercel builds

💡 Pro tip: Use vercel --prebuilt for faster subsequent deployments.

Why SSE for MCP in Serverless?

Traditional TCP MCP servers don't work in serverless environments because they require persistent connections. The HTTP-based MCP Server with Server-Sent Events (SSE) solves this by providing:

Use the SSE endpoint for production deployments on cloud platforms. The TCP server remains available for local development and dedicated server deployments.

📚 Comprehensive deployment guide: See docs/deployment-sse.md for detailed platform-specific instructions, troubleshooting, and performance tuning.

Detailed Setup

Prerequisites

• Python 3.12 | Poetry ≥ 1.8 | Docker + Compose

Environment Variables

Create .env (see .env.example) – minimal:

READWISE_TOKEN=xxxx     # get from readwise.io/api_token
OPENAI_API_KEY=sk-...
DATABASE_URL=postgresql+asyncpg://rw_user:rw_pass@localhost:5432/readwise

All variables are documented in docs/env.md.

Database & Migrations

docker compose up -d db       # Postgres 16 + pgvector
poetry run alembic upgrade head

Sync Commands (CLI)

# first-time full sync
poetry run rwv sync --backfill

# daily incremental (fetch since yesterday)
poetry run rwv sync --since $(date -Idate -d 'yesterday')

Usage Examples

Vector Search (HTTP API)

curl -X POST http://127.0.0.1:8000/search \
     -H 'Content-Type: application/json' \
     -d '{
           "q": "Large Language Models",
           "k": 10,
           "filters": {
             "source": "kindle",
             "tags": ["ai", "research"],
             "highlighted_at": ["2024-01-01", "2024-12-31"]
           }
         }'

Streaming Search (HTTP SSE)

# Real-time streaming via Server-Sent Events (serverless-friendly)
curl -N -H "Accept: text/event-stream" \
  "http://127.0.0.1:8000/mcp/stream?q=neural+networks&k=10"

Streaming Search (MCP TCP)

poetry run python -m readwise_vector_db.mcp --host 0.0.0.0 --port 8375 &

# then from another shell
printf '{"jsonrpc":"2.0","id":1,"method":"search","params":{"q":"neural networks"}}\n' | \
  nc 127.0.0.1 8375

💡 New: Check out the SSE Usage Guide for JavaScript, Python, and browser examples!

Architecture Overview

The system supports multiple deployment patterns to fit different infrastructure needs:

Docker + Local PostgreSQL (Default)

flowchart TB
  subgraph "🐳 Docker Deployment"
    subgraph Ingestion
      A[Readwise API] --> B[Backfill Job]
      C[Nightly Cron] --> D[Incremental Job]
    end
    B --> E[OpenAI Embeddings]
    D --> E
    E --> F[Local PostgreSQL + pgvector]
    F --> G[FastAPI Container]
    G --> H[MCP Server :8375]
    G --> I[Prometheus /metrics]
  end

Vercel + Supabase (Cloud)

flowchart TB
  subgraph Serverless_Deployment
    subgraph Vercel_Edge
      J[FastAPI Serverless]
      K[/health endpoint/]
      L[/search endpoint/]
      M[/docs Swagger UI/]
      J --> K
      J --> L
      J --> M
    end
    subgraph Supabase_Cloud
      N[Managed PostgreSQL]
      O[pgvector Extension]
      P[Automated Backups]
      N --> O
      P --> N
    end
    J -.-> N
    Q[GitHub Actions]
    R[Auto Deploy on Tags]
    Q --> R
    R --> J
  end

Key Differences:

• Docker: Full control, local data, requires infrastructure management

• Vercel + Supabase: Zero-ops, global edge deployment, managed scaling

• Hybrid: Use Supabase with local Docker for development → production consistency

Documentation:

• 📊 Architecture diagrams

• 🚀 SSE deployment patterns

• ⚙️ Platform-specific configurations

Development & Contribution

1. Environment

   poetry install --with dev
   poetry run pre-commit install   # black, isort, ruff, mypy, markdownlint

2. Run tests & coverage

   poetry run coverage run -m pytest && coverage report

3. Performance check (make perf) – fails if /search P95 >500 ms.

4. Branching model: feature/xyz → PR → squash-merge. Use Conventional Commits (feat:, fix: …).

5. Coding style: see .editorconfig and enforced linters.

See CONTRIBUTING.md for full guidelines.

Maintainer Notes

• CI/CD – .github/workflows/ci.yml runs lint, type-check, tests (Py 3.11 + 3.12) and publishes images to GHCR.

• Back-ups – pg_dump weekly cron uploads compressed dump as artifact (Goal G4).

• Releasing – bump version in pyproject.toml, run make release.

License & Credits

Code licensed under the MIT License. Made with ❤️ by the community, powered by FastAPI, SQLModel, pgvector, OpenAI and Taskmaster-AI.