MCP Platform — Retrieval-Augmented Q&A

End-to-end RAG over local PDFs using FAISS, an MCP retrieval server, a FastAPI MCP client, OpenAI, and a small web UI.

Architecture

PDFs in /docs
    │
    ▼
MCP server (:8001/mcp)
  - embed + FAISS index (cached under /data/faiss_index)
  - tools: ensure_index, search_documents, list_documents, index_status
    │  Streamable HTTP (MCP)
    ▼
FastAPI backend (:8000)
  - on boot: ensure_index via MCP
  - LangGraph: retrieve → generate (OpenAI)
  - HTTP API for the UI
    │
    ▼
Frontend (/)

Related MCP server: ragi

Requirements

• Python 3.10+ (3.12 recommended; MCP does not support 3.9)

• An OpenAI API key

Setup

python -m venv .venv

# Windows
.venv\Scripts\activate

# macOS / Linux
source .venv/bin/activate

pip install -r requirements.txt
copy .env.example .env   # or: cp .env.example .env

Put your OpenAI key in .env, then generate sample PDFs (optional):

python scripts/make_sample_pdfs.py

Drop any other .pdf files into docs/.

Run

Start the MCP server first:

python -m mcp_server.server

In a second terminal (venv activated):

python -m backend.main

Open http://127.0.0.1:8000.

API

Notes

• Indexing runs automatically on MCP/backend startup when the PDF fingerprint changes; otherwise the on-disk FAISS index is reused.

• Answers are grounded in retrieved chunks; the model is instructed to say it does not know when context is insufficient.