MCP RAG with Python, LangChain, LangGraph and LlamaIndex

Model Context Protocol (MCP) server project for indexing documents and answering questions with RAG (Retrieval-Augmented Generation). The server uses FastMCP via SSE, LlamaIndex + ChromaDB for retrieval, LangGraph for orchestration and LangChain for generation.

Architecture

Cliente MCP -> SSE (/sse) -> FastMCP
                           -> ingest_documents(paths)
                           -> ask_rag(question)
                                  -> LangGraph
                                     -> recuperar contexto (LlamaIndex/Chroma)
                                     -> gerar resposta (LangChain/OpenAI)
                                     -> validar citações

Related MCP server: Antigravity PDF MCP Server

Quick start

cp .env.example .env
# Edite .env e informe OPENAI_API_KEY
docker compose up --build

The MCP SSE endpoint is at http://localhost:8000/sse.

Documents

To make a local folder available to the container, uncomment the ./docs:/workspace/docs:ro volume in docker-compose.yml. Then call the tool with a path such as /workspace/docs/manual.md.

MCP Tools

ingest_documents

Indexes .txt, .md and .pdf files.

{"paths": ["/workspace/docs/manual.md"]}

ask_rag

Retrieves the most relevant excerpts and answers with [S1], [S2] references.

{"question": "Quais são os procedimentos de backup?", "top_k": 5}

Environment variables

Security

Do not expose SSE publicly without TLS and authentication. Mount only authorized document directories, since the ingestion tool reads the paths provided.