Documentation Fetcher & RAG Search MCP Server

Documentation Fetcher & RAG Search

A modular system for fetching API documentation and enabling semantic search via RAG (Retrieval-Augmented Generation). Designed to give AI coding assistants like Claude access to up-to-date documentation from any project.

Features

• Fetch Documentation: Download complete documentation from API providers in markdown format
• Semantic Search: Hybrid search combining vector embeddings with keyword matching
• MCP Server: Expose search as tools accessible from Claude Code in any project
• Modular Design: Easy to add new documentation sources

Supported Documentation Sources

Quick Start

Prerequisites

• Python 3.12+
• Ollama with bge-m3 model
• Claude Code (for MCP integration)

Installation

# Clone the repository
git clone <repository-url>
cd documentation

# Create virtual environment
python3.12 -m venv .venv
source .venv/bin/activate

# Install dependencies
pip install -r requirements.txt

# Pull the embedding model
ollama pull bge-m3

Fetch & Index Documentation

# Fetch documentation
python -m src.main fetch gemini
python -m src.main fetch fastmcp

# Index for search (requires Ollama running)
python -m src.rag.index gemini
python -m src.rag.index fastmcp

Search Documentation

# Search Gemini docs
python -m src.main search "function calling"

# Search FastMCP docs
python -m src.main search "how to create a tool" -c fastmcp

# More results
python -m src.main search "rate limits" -n 10

MCP Server Integration

The MCP server exposes documentation search as tools that Claude Code can use from any project.

Install in Claude Code

IMPORTANT: MCP configuration requires absolute paths. The cwd field is NOT supported by Claude Code.

Option 1: Using Claude CLI (recommended)

# Replace /path/to/documentation with your actual absolute path
claude mcp add docs-search --scope user --transport stdio -- \
  /path/to/documentation/.venv/bin/python \
  /path/to/documentation/src/mcp_server.py

Option 2: Add to ~/.claude.json manually

{
  "mcpServers": {
    "docs-search": {
      "command": "/path/to/documentation/.venv/bin/python",
      "args": ["/path/to/documentation/src/mcp_server.py"]
    }
  }
}

Common mistakes to avoid:

• Do NOT use cwd - it's not a valid MCP configuration field
• Do NOT use relative paths - they resolve from the caller's directory
• Do NOT use -m src.mcp_server - this requires being in the project directory

Verify Installation

# Check server is registered
claude mcp list

# In Claude Code, check connection status
/mcp

Available Tools

Available Resources

Usage from Claude Code

Once installed, you can ask Claude from any project:

• "Search the gemini docs for function calling"
• "What documentation collections are available?"
• "Search fastmcp for how to create tools"
• "Find rate limit information in gemini docs"

Project Structure

documentation/
├── src/
│   ├── main.py                 # CLI entry point
│   ├── mcp_server.py           # MCP server for Claude Code
│   ├── core/
│   │   ├── fetcher.py          # HTTP/markdown fetching
│   │   └── parser.py           # Navigation parsing
│   ├── modules/
│   │   ├── base.py             # Abstract base class
│   │   ├── gemini/             # Gemini documentation module
│   │   └── fastmcp/            # FastMCP documentation module
│   └── rag/
│       ├── chunker.py          # Markdown-aware chunking
│       ├── embedder.py         # Ollama bge-m3 embeddings
│       ├── sqlite_store.py     # SQLite + sqlite-vec vector store
│       ├── search.py           # Hybrid search with RRF
│       ├── query_expander.py   # Multi-query expansion (LLM)
│       ├── reranker.py         # Cross-encoder reranking
│       └── index.py            # Indexing CLI
├── output/                     # Fetched documentation
│   ├── gemini/
│   └── fastmcp/
├── data/
│   └── docs.db                 # SQLite vector database
├── requirements.txt
└── README.md

Adding New Documentation Sources

1