RAG MCP Server

A Model Context Protocol (MCP) server for Retrieval-Augmented Generation (RAG) operations. This server provides tools for building and querying vector-based knowledge bases from document collections, enabling semantic search and document retrieval capabilities.

• Features
• Architecture
• Installation
• Setup
  • Find the MCP settings file for the client
    • Claude Desktop
    • Claude Code
    • Cursor
    • Cline
    • Windsurf
    • Any other client
  • Set up the MCP server
  • Variant: Manual setup with uvx
• Usage Examples
  • Sample LLM Queries
  • Command Line Examples
• MCP Tools
• Technical Details
• Configuration Examples
• Troubleshooting
• Contributing
• License

Features

• Document Processing: Supports multiple file formats (.txt, .pdf) with automatic text extraction
• Intelligent Chunking: Configurable text chunking with overlap to preserve context
• Vector Embeddings: Uses SentenceTransformers for high-quality text embeddings
• Semantic Search: FAISS-powered similarity search for fast and accurate retrieval
• Incremental Updates: Smart document tracking to only process new or changed files
• Persistent Storage: SQLite-based document store for metadata and change tracking
• Flexible Configuration: Customizable embedding models, chunk sizes, and search parameters

Architecture

rag-mcp-server/
├── src/rag_mcp_server/
│   ├── server.py              # Main MCP server implementation
│   └── core/
│       ├── document_processor.py  # Document loading and chunking
│       ├── embedding_service.py   # Text embedding generation
│       ├── faiss_index.py        # Vector similarity search
│       └── document_store.py     # Document metadata storage

Installation

Using uvx (Recommended)

# Install with uvx (comes with uv)
uvx rag-mcp-server

Using pip

pip install rag-mcp-server

From source

git clone <repository-url>
cd rag-mcp-server
pip install -e .

Setup

The easiest way to run the MCP server is with uvx, but manual setup is also available.

Find the MCP settings file for the client

Claude Desktop

1. Install Claude Desktop as needed
2. Open the config file by opening the Claude Desktop app, going into its Settings, opening the 'Developer' tab, and clicking the 'Edit Config' button
3. Follow the 'Set up the MCP server' steps below

Claude Code

1. Install Claude Code as needed

2. Run the following command to add the RAG server:

   claude mcp add rag
   

   Or manually add with custom configuration:

   claude mcp add-json rag '{"command":"uvx","args":["rag-mcp-server","--knowledge-base","/path/to/your/docs","--embedding-model","all-MiniLM-L6-v2","--chunk-size","1000","--chunk-overlap","200"]}'
   

Cursor

1. Install Cursor as needed
2. Open the config file by opening Cursor, going into 'Cursor Settings' (not the normal VSCode IDE settings), opening the 'MCP' tab, and clicking the 'Add new global MCP server' button
3. Follow the 'Set up the MCP server' steps below

Cline

1. Install Cline in your IDE as needed
2. Open the config file by opening your IDE, opening the Cline sidebar, clicking the 'MCP Servers' icon button that is second from left at the top, opening the 'Installed' tab, and clicking the 'Configure MCP Servers' button
3. Follow the 'Set up the MCP server' steps below

Windsurf

1. Install Windsurf as needed
2. Open the config file by opening Windsurf, going into 'Windsurf Settings' (not the normal VSCode IDE settings), opening the 'Cascade' tab, and clicking the 'View raw config' button in the 'Model Context Protocol (MCP) Servers' section
3. Follow the 'Set up the MCP server' steps below

Any other client

1. Find the MCP settings file, usually something like [client]_mcp_config.json
2. Follow the 'Set up the MCP server' steps below

Set up the MCP server

1. Install uv as needed (uvx comes bundled with uv)

2. Add the following to your MCP setup:

   Basic Configuration:

   {
     "mcpServers": {
       "rag": {
         "command": "uvx",
         "args": ["rag-mcp-server"]
       }
     }
   }
   

   Full Configuration with All Parameters:

   {
     "mcpServers": {
       "rag": {
         "command": "uvx",
         "args": [
           "rag-mcp-server",
           "--knowledge-base", "/path/to/your/documents",
           "--embedding-model", "ibm-granite/granite-embedding-278m-multilingual",
           "--chunk-size", "500",
           "--chunk-overlap", "200",
           "--top-k", "7",
           "--verbose"
         ]
       }
     }
   }
   

Variant: Manual setup with uvx

If you prefer to run the server manually or need specific Python version:

# Run with default settings
uvx rag-mcp-server

# Run with all parameters specified
uvx rag-mcp-server \
  --knowledge-base /path/to/documents \
  --embedding-model "ibm-granite/granite-embedding-278m-multilingual" \
  --chunk-size 500 \
  --chunk-overlap 200 \
  --top-k 7 \
  --verbose

# Run from source directory
uvx --from . rag-mcp-server \
  --knowledge-base /home/user/documents \
  --embedding-model "all-MiniLM-L6-v2" \
  --chunk-size 800 \
  --chunk-overlap 100 \
  --top-k 5

Usage Examples

Sample LLM Queries

Here are example queries you can use with your LLM to interact with the RAG server:

Initialize a knowledge base with custom parameters:

Initialize the knowledge base with:
- knowledge_base_path: "/home/user/research_papers"
- embedding_model: "ibm-granite/granite-embedding-278m-multilingual"
- chunk_size: 300
- chunk_overlap: 50

Search with specific parameters:

Search for "machine learning optimization techniques" in the knowledge base at "/home/user/research_papers" and return the top 10 results with similarity scores.

Initialize with high-quality embeddings:

Set up a knowledge base at "/data/technical_docs" using the "all-mpnet-base-v2" model with chunk_size of 1000 and chunk_overlap of 400 for better context preservation.

Refresh and get statistics:

Refresh the knowledge base at "/home/user/documents" to include any new files, then show me the statistics including total documents, chunks, and current configuration.

List and search documents:

List all documents in the knowledge base, then search for information about "API authentication" and show me the top 5 most relevant chunks.

Complex workflow example:

1. Initialize a knowledge base at "/home/user/project_docs" with embedding_model "all-MiniLM-L6-v2", chunk_size 800, and chunk_overlap 150
2. Show me the statistics
3. Search for "database optimization strategies"
4. List all documents that were processed

Multilingual search example:

Initialize the knowledge base at "/docs/international" using the multilingual model "ibm-granite/granite-embedding-278m-multilingual", then search for "machine learning" in multiple languages and show the top 7 results.

Command Line Examples

High-Quality Configuration for Research:

uvx rag-mcp-server \
  --knowledge-base /home/tommasomariaungetti/RAG \
  --embedding-model "all-mpnet-base-v2" \
  --chunk-size 1000 \
  --chunk-overlap 400 \
  --top-k 10 \
  --verbose

Fast Processing for Large Document Sets:

uvx rag-mcp-server \
  --knowledge-base /data/large_corpus \
  --embedding-model "all-MiniLM-L6-v2" \
  --chunk-size 2000 \
  --chunk-overlap 100 \
  --top-k 5

Multilingual Document Processing:

uvx rag-mcp-server \
  --knowledge-base /docs/multilingual \
  --embedding-model "ibm-granite/granite-embedding-278m-multilingual" \
  --chunk-size 500 \
  --chunk-overlap 200 \
  --top-k 7

Running from Source with Custom Settings:

uvx --from . rag-mcp-server \
  --embedding-model "all-MiniLM-L6-v2" \
  --chunk-size 800 \
  --chunk-overlap 100 \
  --top-k 5 \
  --knowledge-base /home/tommasomariaungetti/RAG

MCP Tools

The following tools are available:

1. initialize_knowledge_base

Initialize a knowledge base from a directory of documents.

Parameters:

• knowledge_base_path (optional): Path to document directory - defaults to server config
• embedding_model (optional): Model name for embeddings - defaults to "ibm-granite/granite-embedding-278m-multilingual"
• chunk_size (optional): Maximum chunk size in characters - defaults to 500
• chunk_overlap (optional): Chunk overlap size in characters - defaults to 200

Example Tool Call:

{
  "tool": "initialize_knowledge_base",
  "arguments": {
    "knowledge_base_path": "/path/to/docs",
    "embedding_model": "all-mpnet-base-v2",
    "chunk_size": 1000,
    "chunk_overlap": 200
  }
}

Example LLM Query:

"Initialize a knowledge base from /home/user/documents using the all-mpnet-base-v2 embedding model with 1000 character chunks and 200 character overlap"

2. semantic_search

Perform semantic search on the knowledge base.

Parameters:

• query: Search query text
• knowledge_base_path (optional): Path to knowledge base - defaults to current KB
• top_k (optional): Number of results to return - defaults to 7
• include_scores (optional): Include similarity scores - defaults to false

Example Tool Call:

{
  "tool": "semantic_search",
  "arguments": {
    "query": "How to implement RAG systems?",
    "knowledge_base_path": "/path/to/docs",
    "top_k": 5,
    "include_scores": true
  }
}

Example LLM Query:

"Search for 'machine learning optimization techniques' and show me the top 5 results with similarity scores"

3. refresh_knowledge_base

Update the knowledge base with new or changed documents.

Parameters:

• knowledge_base_path (optional): Path to knowledge base - defaults to current KB

Example Tool Call:

{
  "tool": "refresh_knowledge_base",
  "arguments": {
    "knowledge_base_path": "/path/to/docs"
  }
}

Example LLM Query:

"Refresh the knowledge base to include any new or modified documents"

4. get_knowledge_base_stats

Get detailed statistics about the knowledge base.

Parameters:

• knowledge_base_path (optional): Path to knowledge base - defaults to current KB

Example Tool Call:

{
  "tool": "get_knowledge_base_stats",
  "arguments": {
    "knowledge_base_path": "/path/to/docs"
  }
}

Example LLM Query:

"Show me the statistics for the knowledge base including document count, chunk information, and current configuration"

5. list_documents

List all documents in the knowledge base with metadata.

Parameters:

• knowledge_base_path (optional): Path to knowledge base - defaults to current KB

Example Tool Call:

{
  "tool": "list_documents",
  "arguments": {
    "knowledge_base_path": "/path/to/docs"
  }
}

Example LLM Query:

"List all documents in the knowledge base with their chunk counts and metadata"

Technical Details

Document Processing

The system uses a sophisticated document processing pipeline:

1. File Detection: Scans directories for supported file types
2. Content Extraction:
   • Plain text files: Direct UTF-8/Latin-1 reading
   • PDF files: PyMuPDF-based text extraction
3. Text Chunking:
   • Splits documents into manageable chunks
   • Preserves word boundaries
   • Maintains context with configurable overlap

Embedding Generation

• Default Model: ibm-granite/granite-embedding-278m-multilingual
• Batch Processing: Efficient batch encoding for large document sets
• Fallback Support: Automatic fallback to all-MiniLM-L6-v2 if primary model fails
• Progress Tracking: Visual progress bars for large operations

Vector Search

• Index Type: FAISS IndexFlatIP (Inner Product)
• Similarity Metric: Cosine similarity (via L2 normalization)
• Performance: Scales to millions of documents
• Accuracy: Exact nearest neighbor search

Document Store

• Storage: SQLite database
• Tracking: File hash, modification time, chunk count
• Incremental Updates: Only processes changed files
• Location: Stored alongside knowledge base documents

Configuration Examples

MCP Client Configurations

Basic Configuration (Claude Desktop/Cursor/Cline):

{
  "mcpServers": {
    "rag": {
      "command": "uvx",
      "args": ["rag-mcp-server"]
    }
  }
}

Full Configuration with All Parameters:

{
  "mcpServers": {
    "rag": {
      "command": "uvx",
      "args": [
        "rag-mcp-server",
        "--knowledge-base", "/path/to/documents",
        "--embedding-model", "ibm-granite/granite-embedding-278m-multilingual",
        "--chunk-size", "500",
        "--chunk-overlap", "200",
        "--top-k", "7",
        "--verbose"
      ]
    }
  }
}

Multiple Knowledge Base Configuration:

{
  "mcpServers": {
    "rag-technical": {
      "command": "uvx",
      "args": [
        "rag-mcp-server",
        "--knowledge-base", "/docs/technical",
        "--embedding-model", "all-mpnet-base-v2",
        "--chunk-size", "1000",
        "--chunk-overlap", "400"
      ]
    },
    "rag-research": {
      "command": "uvx",
      "args": [
        "rag-mcp-server",
        "--knowledge-base", "/docs/research",
        "--embedding-model", "all-MiniLM-L6-v2",
        "--chunk-size", "500",
        "--chunk-overlap", "100",
        "--port", "8001"
      ]
    }
  }
}

Command Line Examples

High-Quality Configuration for Research:

uvx rag-mcp-server \
  --knowledge-base /path/to/research/docs \
  --embedding-model "all-mpnet-base-v2" \
  --chunk-size 1000 \
  --chunk-overlap 400 \
  --top-k 10

Fast Processing Configuration:

uvx rag-mcp-server \
  --knowledge-base /path/to/large/corpus \
  --embedding-model "all-MiniLM-L6-v2" \
  --chunk-size 2000 \
  --chunk-overlap 100 \
  --top-k 5

Multilingual Configuration:

uvx rag-mcp-server \
  --knowledge-base /path/to/multilingual/docs \
  --embedding-model "ibm-granite/granite-embedding-278m-multilingual" \
  --chunk-size 500 \
  --chunk-overlap 200 \
  --top-k 7

Development Configuration with Verbose Logging:

uvx --from . rag-mcp-server \
  --knowledge-base ./test_documents \
  --embedding-model "all-MiniLM-L6-v2" \
  --chunk-size 300 \
  --chunk-overlap 50 \
  --top-k 3 \
  --verbose

Error Handling

The server implements comprehensive error handling:

• File Access Errors: Graceful handling of permission issues
• Encoding Errors: Automatic encoding detection and fallback
• Model Loading Errors: Fallback to default models
• Database Errors: Transaction rollback and recovery
• Search Errors: Informative error messages

Performance Considerations

Memory Usage

• Embeddings are stored in memory for fast search
• Approximate memory: num_chunks × embedding_dimension × 4 bytes
• Example: 10,000 chunks × 384 dimensions ≈ 15 MB

Processing Speed

• Document processing: ~100-500 docs/minute (depending on size)
• Embedding generation: ~50-200 chunks/second (model dependent)
• Search latency: <10ms for 100K documents

Optimization Tips

1. Use smaller embedding models for faster processing
2. Increase chunk size for fewer chunks (may reduce accuracy)
3. Decrease overlap for faster processing (may lose context)
4. Use SSD storage for document store database

Development

Running Tests

pytest tests/

Code Formatting

black src/
isort src/

Type Checking

mypy src/

Troubleshooting

Common Issues

1. "No knowledge base path provided"

   • Solution: Either provide path in tool call or use --knowledge-base flag

2. "Model mismatch detected"

   • Solution: This is a warning; the system will use the closest available model

3. "Failed to initialize embedding model"

   • Solution: Check internet connection or use a locally cached model

4. "No documents found in knowledge base"

   • Solution: Ensure directory contains .txt or .pdf files

Debug Mode

Enable verbose logging for troubleshooting:

uvx rag-mcp-server --verbose

Help and Resources

• GitHub Repository
• PyPI Package
• MCP Documentation
• Issue Tracker
• Email: your-email@example.com

Contributing

Contributions are welcome! Please:

1. Fork the repository
2. Create a feature branch
3. Add tests for new functionality
4. Ensure all tests pass
5. Submit a pull request

License

MIT License - see LICENSE file for details.

Acknowledgments

• Built on MCP (Model Context Protocol)
• Powered by Sentence Transformers
• Vector search by FAISS
• PDF processing by PyMuPDF