RAG MCP: Document Processing Server

A Retrieval-Augmented Generation (RAG) server built on the Model Context Protocol (MCP) for intelligent document processing and question answering.

Overview

RAG MCP is a tool that allows you to index various document formats and perform semantic searches against them. It uses advanced embedding techniques and vector databases to make your documents searchable through natural language queries.

Features

• Document Indexing: Support for various document formats (PDF, DOCX, XLSX, PPTX, Markdown, AsciiDoc, HTML, XHTML, CSV)
• Semantic Search: Query your documents using natural language
• Flexible Embedding Models: Choose between HuggingFace BGE (default) or Ollama embeddings.
• High Performance: Optimized for various hardware configurations with automatic device selection (CUDA, MPS, CPU) for HuggingFace embeddings.
• Persistent Storage: Vector embeddings are stored locally for future use

Requirements

• Python 3.11+
• Environment with access to your documents
• (Optional) Ollama installed and running if using Ollama embeddings.

Installation

1. Install UV

First, you need to install UV, a Python package installer and resolver:

On macOS/Linux:

curl -sSf https://astral.sh/uv/install.sh | sh

On Windows:

powershell -c "irm https://astral.sh/uv/install.ps1 | iex"

2. Run RAG MCP

Once UV is installed, you can run RAG MCP directly using:

uvx rag-mcp

This will start the MCP server and make it available for document processing.

Environment Variables:

You can configure RAG MCP using environment variables:

• PERSIST_DIRECTORY (Required): Path to the directory where the vector database will be stored (e.g., /path/to/your/persist/directory). A chromadb subfolder will be created here.
• USE_OLLAMA_EMBEDDING (Optional): Set to True to use Ollama embeddings instead of the default HuggingFace BGE embeddings. Requires Ollama to be running. Defaults to False.
• OLLAMA_EMBEDDING_MODEL (Optional): Specifies the Ollama model name to use for embeddings (e.g., nomic-embed-text). Only used if USE_OLLAMA_EMBEDDING=True. Defaults to bge-m3:latest.

Example:

PERSIST_DIRECTORY=/data/rag_db USE_OLLAMA_EMBEDDING=True OLLAMA_EMBEDDING_MODEL=nomic-embed-text uvx rag-mcp

IDE Integration

VS Code Integration

To integrate with Visual Studio Code, create a mcp.json file with the following content:

{
    "servers": {
        "rag-mcp-server": {
            "type": "stdio",
            "command": "uvx",
            "args": [
                "rag-mcp"
            ],
            "env": {
                "PERSIST_DIRECTORY": "/path/to/your/persist/directory",
                // Optional: Uncomment and set to true to use Ollama
                // "USE_OLLAMA_EMBEDDING": "True",
                // Optional: Specify Ollama model if using Ollama
                // "OLLAMA_EMBEDDING_MODEL": "nomic-embed-text"
            }
        }
    }
}

Replace /path/to/your/persist/directory with the desired path. Adjust Ollama settings as needed.

Cursor Integration

To integrate with Cursor, go to Cursor Settings > MCP and paste this configuration:

{
    "mcpServers": {
      "rag-mcp": {
        "command": "uvx",
        "args": ["rag-mcp"],
        "env": {
          "PERSIST_DIRECTORY": "/path/to/your/persist/directory",
          // Optional: Uncomment and set to true to use Ollama
          // "USE_OLLAMA_EMBEDDING": "True",
          // Optional: Specify Ollama model if using Ollama
          // "OLLAMA_EMBEDDING_MODEL": "nomic-embed-text"
        }
      }
    }
  }

Replace /path/to/your/persist/directory with the desired path. Adjust Ollama settings as needed.

Usage

Supported Embedding Models

The system supports the following embedding models:

1. HuggingFace BGE Embeddings (default): High-quality embeddings that work offline. Optimized for available hardware (CUDA, MPS, CPU).
   • Uses BAAI/bge-m3 model.
2. Ollama Embeddings (optional): Uses embeddings generated by a running Ollama instance. Enable by setting USE_OLLAMA_EMBEDDING=True.
   • Default model: bge-m3:latest.
   • Specify a different model using OLLAMA_EMBEDDING_MODEL.

How It Works

1. Document Loading: Uses DoclingLoader to parse various document formats
2. Text Splitting: Documents are split into manageable chunks using RecursiveCharacterTextSplitter
3. Embedding Generation: Text chunks are converted to vector embeddings using either HuggingFace BGE or Ollama.
4. Storage: Embeddings are stored in a Chroma vector database
5. Retrieval: When queried, the system finds semantically similar content to answer questions

Acknowledgments

• This project uses LangChain and Docling for intelligent document parsing
• Vector storage provided by Chroma
• Embedding models from HuggingFace and potentially Ollama
• Built on the Model Context Protocol (MCP)