fayna-rag-mcp

{"type": "text"}# fayna-rag-mcp — local knowledge base with RAG and MCP

Developed by Fayna Digital Author: Volodymyr Shevchenko

Problem: the team collects documentation (policies, manuals, notes) in files and quickly loses natural-language access to it — searching by file name or Ctrl+F does not scale, and sending internal documents to a cloud LLM service is not always acceptable for privacy reasons.

Solution: a local RAG pipeline (FAISS + multilingual embeddings) based on a local LLM (Ollama) — and the same search/Q&A exposed as an MCP server, so any MCP client (Claude Desktop/Code, etc.) or an external automated workflow can use it via simple REST routes. No data leaves the machine it runs on.

Result: five MCP tools and four REST routes for semantic search, RAG Q&A with sources, document reading, and offline cataloging — ready to connect to Claude or n8n in minutes, with no cloud dependencies.

Features

Related MCP server: OpenLMlib

Stack

Python 3.10+ · FAISS (faiss-cpu) · sentence-transformers · tiktoken · Ollama · FastMCP · Tesseract/poppler/Whisper for multi-format ingest · Docker.

RAG Pipeline

docs/ → load (.txt/.md/.pdf/.docx) → chunk (tiktoken) → embed (mpnet) → FAISS → retrieve → Ollama → answer + sources

• Chunking — by tiktoken tokens (cl100k_base), not characters. Default CHUNK_SIZE=700 tokens, CHUNK_OVERLAP=100 tokens.

• Embeddings: paraphrase-multilingual-mpnet-base-v2 — a multilingual model (UA/PL/EN/RU and others) so search works regardless of query language.

• LLM: any Ollama model, default qwen2.5:7b.

• Index: FAISS IndexFlatIP (cosine similarity), TOP_K=5.

Quick start

pip install -r src/requirements.txt

# Przykład: demo-korpus na kilka dokumentów (sample-docs/)
export DOCUMENTS_DIR=./sample-docs
python -m src.main build-index      # → src/index/index.faiss + chunks.pkl

# Interaktywne Q&A (CLI)
python -m src.main

# Serwer MCP (transport z env MCP_TRANSPORT: stdio|http)
python -m src.mcp.server

Tests:

pip install -r tests/requirements-dev.txt
pytest -q

Docker

docker compose up -d

Default values in docker-compose.yml/Dockerfile are designed for Ollama running on the host (via host.docker.internal); adjust OLLAMA_URL to your own network (bridge address on Linux, separate Ollama container, etc.).

Configuration (src/config.py, everything via env)

Connecting to an MCP client (e.g., Claude Code) — via its MCP server configuration, with the command python -m src.mcp.server (stdio) or the container URL (http).

Structure

src/
├── config.py         # wszystkie env-zmienne + domyślne
├── main.py           # CLI: build-index | interaktywne Q&A
├── assistant.py       # CompanyKBAssistant (LLM decyduje czy wołać MCP-toolki)
├── catalog.py         # offline-klasyfikacja dokumentów przez Ollama → JSON+HTML
├── ingest.py           # multi-formatowy ingest: OCR skanów, vision-opis diagramów, Whisper-transkrypcja
├── rag/
│   ├── ingest.py      # load_document (.txt/.md/.pdf/.docx)
│   ├── chunk.py        # chunk_text (tiktoken cl100k_base, overlap)
│   ├── embed.py         # embed_chunks (sentence-transformers)
│   ├── build_index.py   # build_index → FAISS + pickle
│   └── query.py          # retrieve / build_prompt / ask
└── mcp/
    ├── server.py     # FastMCP: 5 MCP-toolków + 4 trasy REST
    └── client.py      # MCPClient (JSON-RPC przez subprocess)

Multi-format ingest (OCR of scans via Tesseract, description of drawings/diagrams via a vision model, audio/video transcription via Whisper) — a separate, dependency-heavier path, not needed for the basic text corpus above.

License

MIT — see LICENSE. © Fayna Digital.