mcp-rag: Semantic RAG served over the Model Context Protocol

A semantic Retrieval-Augmented Generation engine, exposed as an MCP server so any MCP client (Claude Desktop, an IDE agent, …) can search and question a knowledge base as a native tool.

 documents ─▶ chunk ─▶ embed ─▶ index ─┐
                                        ├─▶ hybrid retrieval ─▶ grounded answer
 query ─────────────────────────────────┘   (semantic + BM25,     with citations
                                              fused by RRF)              │
                                                                         ▼
                                                    MCP server ─▶ any MCP client

The whole thing runs offline out of the box, local LSA embeddings + an extractive, citation-grounded answerer, with zero API keys or model downloads. Production backends (Voyage / OpenAI / sentence-transformers for embeddings, Anthropic / OpenAI for generation) are a one-line config switch.

Demo corpus: a self-contained fictional SaaS knowledge base ("Nimbus", a cloud data platform): authentication, billing, rate limits, data retention, security, incident runbook, webhooks, SDK. Nothing copyrighted; every answer is traceable to a source.

What it does

Ask a question phrased in your own words and get an answer grounded in the docs:

$ python scripts/demo_query.py "what happens if I go over my included usage?"

A: When you exceed your included quota, Nimbus does not cut off your service;
   instead, additional usage is billed as overage at the metered rate
   ($0.50 per extra 10k calls, $0.10 per extra GB). [1]
Sources:
   [1] billing.md, Billing and quotas

Note there's no keyword overlap between "go over my included usage" and "exceed your quota / overage", that match is semantic, which is the point.

Related MCP server: rag-mcp

MCP tools exposed

Verified end-to-end over the real MCP stdio protocol (see tests/). Register it in a client with mcp.json.

Retrieval ablation (computed by make eval)

17 gold questions, deliberately paraphrased away from the documents' wording. hit@k = correct document in the top-k; MRR = how highly it's ranked.

All three usually find the right document on this clean corpus, but hybrid ranks it highest most consistently, fusing dense (semantic) and sparse (keyword) retrieval is a tuning-free win, and the paraphrased questions are exactly where pure keyword search ranks worse.

With a neural embedding backend (Voyage/OpenAI/ST) on a larger, noisier corpus the gap between lexical and semantic widens further; the local LSA backend keeps the demo runnable anywhere while preserving the same ranking behaviour.

Quickstart

pip install -r requirements.txt && pip install -e .

make demo     # ask a question from the CLI
make eval     # retrieval ablation → reports/eval_results.json
make server   # run the MCP server (stdio)
make test     # 10 tests, incl. an end-to-end MCP protocol check

Use it from Claude Desktop

Copy mcp.json into your client config (set the absolute cwd), restart the client, and the four tools appear. Ask "search the Nimbus docs for how failover works" and the model calls search_documents / answer_question.

Switch to production backends

pip install -r requirements-prod.txt
export MCPRAG_EMBEDDING_BACKEND=voyage   VOYAGE_API_KEY=...
export MCPRAG_GENERATOR_BACKEND=anthropic ANTHROPIC_API_KEY=...

No code changes, the retriever and server are backend-agnostic.

Layout

src/mcprag/
  ingest.py            markdown loading + section-aware chunking w/ overlap
  embeddings/          Embedder protocol · local LSA (offline) · neural backends
  index/vector_store.py  cosine + BM25 + hybrid RRF retrieval
  generator/           extractive (offline, cited) · LLM backends
  rag.py               RAGEngine (ingest→embed→index→retrieve→generate)
  evaluation.py        hit@k / MRR
  server.py            FastMCP server exposing the tools
data/corpus/           the Nimbus knowledge base (8 markdown docs)
eval/qa_gold.json      paraphrased gold questions
scripts/               demo_query · run_eval
tests/                 retrieval, answers, evaluation, MCP protocol

Design notes

• Grounded by construction. The offline generator only emits sentences taken verbatim from retrieved chunks, each with a citation, it cannot hallucinate.

• Hybrid retrieval. Reciprocal Rank Fusion of dense + sparse rankings needs no weight tuning and is robust across query types.

• Backend-agnostic. Embeddings and generation are pluggable Protocols; offline and production share identical retrieval/serving code.

See docs/ARCHITECTURE.md, docs/RESULTS.md, and docs/IMPROVEMENTS.md.

Tech stack

Python · MCP SDK (FastMCP) · scikit-learn (TF-IDF + LSA) · rank_bm25 · numpy · pydantic. Optional: Voyage / OpenAI / sentence-transformers / Anthropic.

License

MIT. The demo corpus is fictional.