A Model Context Protocol (MCP) server for searching OpenMM molecular dynamics simulation documentation. This server vectorizes documentation and provides semantic search capabilities optimized for integration with large language models (LLMs).

🌐 Language/言語: English | 日本語

This MCP server allows natural language search through OpenMM documentation. It encodes documentation content into vector embeddings using modern language models and stores them in a FAISS vector database for efficient retrieval. When a query is received, the server finds the most semantically relevant documentation sections and returns them, making it particularly useful for:

• Finding relevant documentation on molecular dynamics methods
• Searching for usage instructions of OpenMM functions and classes
• Retrieving information about simulation parameters and settings
• Getting code examples and implementation details

• Semantic Search: Find documentation based on meaning, not just keywords
• MCP Integration: Fully compatible with Claude Desktop and other MCP-enabled applications
• Multi-language Support: Process queries in both English and Japanese
• Efficient Retrieval: Uses FAISS for high-performance vector similarity search
• Customizable: Configurable embedding models and search parameters

• Python 3.9 or higher
• uv package manager (recommended) or pip
• Minimum 8GB RAM (16GB+ recommended for index creation)

# Clone the repository
git clone https://github.com/yourusername/openMM-Doc-MCP.git
cd openMM-Doc-MCP

# Create and activate virtual environment using uv
uv venv

# Install dependencies with uv (recommended)
uv pip install -r requirements.txt

# Or, if using pip
# python -m venv .venv
# source .venv/bin/activate  # For Linux/macOS
# .venv\Scripts\activate     # For Windows
# pip install -r requirements.txt

Before using the server, you need to create a vector index of the OpenMM documentation:

uv run python create_faiss_index.py

Optional parameters:

# Specify a custom document directory
uv run python create_faiss_index.py --docs_dir "/path/to/docs"

# Specify a custom output directory
uv run python create_faiss_index.py --output_dir "/path/to/output"

# Specify a custom embedding model
uv run python create_faiss_index.py --embedding_model "intfloat/multilingual-e5-large"

Index creation may take some time and requires significant memory. By default, index files are created in the data/indices/docs/ directory.

You can configure the server using environment variables:

# Set server port (default is 8080)
export MCP_SERVER_PORT=8888

# Set index directory (optional)
export MCP_INDEX_DIR="/path/to/custom/index"

For Windows:

set MCP_SERVER_PORT=8888
set MCP_INDEX_DIR=C:\path\to\custom\index

uv run python server.py

By default, the server listens at http://localhost:8080.

You can directly search from the command line:

uv run python search_molecular_simulation.py "how to set up a water box simulation"

You can also send HTTP requests to the server:

POST http://localhost:8080/query
Content-Type: application/json

{
  "query": "how to set up a water box simulation",
  "top_k": 5
}

Edit the Claude Desktop configuration file to add this MCP server. The configuration file path is:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

Add the following JSON configuration (within the existing mcpServers object):

{
  "mcpServers": {
    "OpenMM Documentation": {
      "command": "uv",
      "args": [
        "run",
        "--with",
        "mcp[cli]",
        "--with",
        "faiss-cpu",
        "--with",
        "langchain",
        "--with",
        "sentence-transformers",
        "mcp",
        "run",
        "/path/to/openMM-Doc-MCP/server.py"
      ]
    }
  }
}

Notes:

• If the uv command is not in your environment path, use an absolute path (e.g.: /path/to/uv).
• Replace /path/to/openMM-Doc-MCP/server.py with the absolute path to this script.
• Always use absolute paths, not relative paths.

If Claude Desktop cannot connect to the MCP server:

1. Verify that the path to server.py in the configuration file is correct (absolute path)
2. Make sure uv is properly installed and accessible
3. Check system logs for any errors
4. Try restarting Claude Desktop after making changes to the configuration

This server implements the Model Context Protocol (MCP) and provides the following tools:

1. search_documents

   • Searches for similar documents based on a query string
   • Parameters:
     • query: The search query text (required)
     • top_k: Number of results to return (default 5)
     • index_path: Optional custom path to FAISS index
   • Returns: Dictionary containing search results with relevant document sections

2. get_index_info

   • Gets information about the currently loaded vector database index
   • Parameters:
     • index_path: Optional custom path to FAISS index
   • Returns: Dictionary containing index information

For detailed API specifications, see:

• specs/apispec_en.md
• specs/apispec_ja.md

This project includes tests for the vector database and MCP server.

# Run all tests
uv run -m pytest

# Run specific tests
uv run -m pytest tests/test_server.py
uv run -m pytest src/vector_db/tests/

# Verbose output
uv run -m pytest tests/test_server.py -v

1. Server Won't Start

   • Check that dependencies are correctly installed
   • Verify that the port is not already in use
   • Check logs for specific error messages

2. No Search Results Returned

   • Verify that index files have been created correctly
   • Ensure that the query is not empty
   • Check that the index path is correct

3. Memory Errors

   • Index creation requires sufficient memory (16GB+ recommended)
   • If using large models, try switching to smaller ones

To enable detailed logging:

export DEBUG=true
uv run python server.py

1. NumPy Errors

   • If you encounter errors with libgfortran.5.dylib, running tests with uv run can resolve this
   • In some environments, you may need to install the library with conda install -c conda-forge libgfortran

2. Tests Timeout

   • Increase test timeout: uv run -m pytest --timeout=30

• Use smaller embedding models for faster inference
• Enable GPU acceleration with faiss-gpu (if you have compatible GPUs)
• Adjust chunk size and overlap parameters during indexing

For processing large document collections:

# Adjust chunk size in create_faiss_index.py
uv run python create_faiss_index.py --chunk_size 256 --chunk_overlap 20

openMM-Doc-MCP/
├── create_faiss_index.py   # Index creation script
├── search_molecular_simulation.py # CLI search utility
├── server.py               # MCP server implementation
├── data/
│   └── indices/
│       └── docs/           # Default location for index files
│           ├── index.faiss # FAISS index file
│           └── index.pkl   # Metadata pickle file
├── specs/
│   ├── apispec_en.md       # API specification (English)
│   └── apispec_ja.md       # API specification (Japanese)
├── src/
│   └── vector_db/          # Vector database related modules
│       ├── indexer.py      # Indexer implementation
│       ├── retriever.py    # Retriever implementation
│       └── tests/          # Vector DB tests
└── tests/
    └── test_server.py      # Server tests

This project is provided under the LICENSE.