graphrag-ts citation provenance

graphrag-ts

Attribution is stored at write time, not asserted at read time.

graphrag-ts is a GraphRAG engine in TypeScript. It builds the knowledge graph in PostgreSQL, keeps source evidence on every node and edge, and answers with retrievable citations — or returns an explicit abstention when the coverage behind an answer is too thin to support it.

Benchmarks / 基准报告

English report · 中文报告

Hosted endpoint / 托管端点

https://mcp.ashesborn.cloud/mcp (anonymous tier: free integrity_paper_status, the two judgments answer with an x402 payment requirement; a keyed session unlocks them) · free DOI status: https://pay.ashesborn.cloud/api/v1/judge/paper?doi= · x402 wire format

Try it in 60 seconds

No install, no API key, no model endpoint. Two ways in:

• In a browser: Retraction Status — paste a DOI, get the source-by-source retraction status and the evidence tiers behind it, with no local model and no account.
• From the command line:

curl -s 'https://pay.ashesborn.cloud/api/v1/judge/paper?doi=10.1016/S0140-6736(97)11096-0'

The response is the paper plus every retraction record behind it, each with its source_tier and source_uri — the evidence the graph stores, not a model's recollection of it.

Point any Streamable HTTP MCP client at the hosted endpoint (public read tier, 20 requests/minute per IP):

{

"mcpServers": { "graphrag-ts": { "url": "https://mcp.ashesborn.cloud/mcp" } } }

A bearer API key unlocks the full tool set (graph queries, index jobs, memory writes). The two judgment tools also answer per call without a key via x402 payment ($0.02 / $0.10).

What it does differently

• Postgres-native. Graph tables and pgvector live in the database you already operate — no Neo4j sidecar and no second store to keep in sync.
• Incremental by construction. Rebuilds reuse unchanged community summaries, so the cost of a refresh tracks the delta rather than the corpus size.
• Two retrieval shapes. Local: hybrid vector, keyword and graph retrieval over a community's evidence. Global: Map-Reduce over community summaries for corpus-wide questions.
• Explicit abstention. Judgment verdicts carry source tiers and return an abstention when reference coverage is thin, instead of answering from a weak signal.
• Evidence at chunk level. Entities and relationships keep the document and chunk they were extracted from, so an answer can be traced back to the text that produced it.

MCP tools

Quick start — self-hosted

MCP with Docker Compose

The Compose stack starts PostgreSQL/pgvector and the MCP service. You must provide reachable chat and embedding endpoints; see the MCP guide for configuration, client setup, and workspace behavior.

cp .env.example .env

Configure the database password and model endpoints in .env.

docker compose up -d --build curl http://127.0.0.1:3000/healthz

Connect an MCP client to http://127.0.0.1:3000/mcp. For the optional local browser showcase, configure the models and run docker compose --profile showcase up -d --build --wait, then open http://127.0.0.1:8080. The showcase is a local demo, not an authentication layer; do not expose it to untrusted users.

Use the library

pnpm install

pnpm run db:generate cp .env.example .env

Configure DATABASE_URL and RAG_* model settings; see the guides below.

pnpm run db:push bun run examples/demo.ts

The library requires Bun, PostgreSQL with pgvector, and compatible chat and embedding endpoints. See the English guide or 中文指南 for environment setup, API examples, adapters, and tuning.

Overview

• Markdown ingestion, entity/relationship extraction, and community detection
• Full and incremental builds, with unchanged community summaries reused
• Hybrid vector, keyword, and graph retrieval, plus corpus-wide Map-Reduce queries
• MCP server with isolated workspaces and a local browser showcase

See the architecture guide for the pipeline and module details.

Benchmark

The reported in-corpus run reached 10/12 strict hits and 95.8% combined lexical recall on 12 questions from five indexed chapters. This is a small project benchmark, not human semantic grading or a cross-project comparison. The reports include methodology, per-query results, and limitations.

The English and Chinese reports are linked at the top of this page. To run the three-question Docker smoke benchmark, use pnpm benchmark:docker:quick with the model endpoints configured and loaded. See the MCP guide for prerequisites and the report for benchmark scope.

Listed on

• Official MCP Registry — io.github.sadofriod/graphrag-ts
• Glama — server rating, plus the connector listing for the hosted endpoint
• mcpservers.org

Documentation

• docs/architecture.md: architecture and data flow
• docs/en-US.md: English user guide
• docs/zh-CN.md: Chinese user guide
• docs/comparison.md: comparison notes
• integrations/mcp/docs/x402-paid-api.md: optional x402 paid query API, dataset bundles
• integrations/mcp/docs/hosted-beta.md: hosted pilot operations checklist
• integrations/mcp/docs/x402-discovery.md: how agents discover the paid API (Bazaar, .well-known/x402)
• integrations/mcp/docs/deploy-x402.md: deploying the paid API on a Docker/Colima host behind Cloudflare Tunnel
• CONTRIBUTING.md: contribution guide

Contributing

Contributions are welcome through issues and pull requests. The project is organized around small, testable modules, so the most useful changes usually fit into one of these areas:

• graph construction quality
• retrieval quality and ranking
• namespace isolation
• model-loading robustness
• documentation and examples

License

MIT