Epstein Files RAG MCP Server

A Model Context Protocol (MCP) server that provides Retrieval-Augmented Generation (RAG) capabilities over the Epstein Files dataset from HuggingFace.

Features

• 🔍 Semantic search over 20K+ Epstein Files documents
• 🚀 Runs entirely on CPU and RAM
• 💾 Vector storage on NVME via Qdrant Docker
• 🎯 Uses all-MiniLM-L6-v2 embedding model
• 📦 Zero local files needed - run directly via uv

Prerequisites

1. Docker - for running Qdrant
2. UV - Python package manager

Install UV:

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

Quick Start

1. Start Qdrant Docker Container

# Create storage directory on your NVME
mkdir -p /path/to/nvme/qdrant_storage

# Run Qdrant
docker run -d \
  --name qdrant \
  -p 6333:6333 \
  -p 6334:6334 \
  -v /path/to/nvme/qdrant_storage:/qdrant/storage \
  qdrant/qdrant

2. Configure Your LLM Client

Add this to your MCP servers configuration (e.g., Claude Desktop config):

{
  "mcpServers": {
    "epstein-rag": {
      "command": "uvx",
      "args": [
        "--from",
        "git+https://github.com/justinlime/epstein-rag-mcp.git",
        "epstein-rag-mcp"
      ],
      "env": {
        "QDRANT_HOST": "localhost",
        "QDRANT_PORT": "6333"
      }
    }
  }
}

3. Restart Your LLM Client

That's it! The server will automatically:

• Download and install dependencies
• Load the embedding model
• Fetch the dataset from HuggingFace (first run only)
• Create embeddings and index them in Qdrant
• Be ready to answer queries

Usage

Once configured, your LLM can use the search_epstein_files tool:

Example queries:

• "Find documents mentioning flight logs"
• "Search for references to specific individuals"
• "What documents discuss financial transactions?"

Configuration

Environment Variables

• QDRANT_HOST - Qdrant server host (default: localhost)
• QDRANT_PORT - Qdrant server port (default: 6333)

First Run

The first time you run the server, it will:

1. Download the all-MiniLM-L6-v2 model (~80MB)
2. Fetch the Epstein Files dataset from HuggingFace (~100MB)
3. Generate embeddings for all documents (10-30 minutes on CPU)
4. Index them in Qdrant

Subsequent runs will be instant as the data is persisted in Qdrant.

Development

Local Installation

git clone https://github.com/yourusername/epstein-rag-mcp.git
cd epstein-rag-mcp
uv pip install -e .

Running Locally

uv run epstein-rag-mcp

Docker Management

# View logs
docker logs qdrant

# Stop Qdrant
docker stop qdrant

# Start Qdrant
docker start qdrant

# Access web UI
# Open http://localhost:6333/dashboard

Troubleshooting

"Failed to connect to Qdrant"

Make sure the Docker container is running:

docker ps | grep qdrant

Port Already in Use

Change the port mapping:

docker run -d --name qdrant -p 6335:6333 ...

Then update QDRANT_PORT to 6335 in your config.

Slow Indexing

This is normal on CPU. The first run can take 10-30 minutes depending on your hardware. Reduce BATCH_SIZE in the code if you run out of memory.

Architecture

┌─────────────┐
│  LLM Client │
└──────┬──────┘
       │ MCP Protocol
       │
┌──────▼──────────────────┐
│  epstein-rag-mcp        │
│  ┌──────────────────┐   │
│  │ Sentence         │   │
│  │ Transformer      │   │
│  │ (all-MiniLM-L6)  │   │
│  └─────────┬────────┘   │
│            │             │
│  ┌─────────▼────────┐   │
│  │ Dataset Loader   │   │
│  │ (HuggingFace)    │   │
│  └──────────────────┘   │
└────────┬─────────────────┘
         │
    ┌────▼─────┐
    │  Qdrant  │
    │  Docker  │
    └────┬─────┘
         │
    ┌────▼─────┐
    │   NVME   │
    └──────────┘

License

MIT

Contributing

Contributions welcome! Please open an issue or PR.

Credits

• Dataset: tensonaut/EPSTEIN_FILES_20K
• Embedding Model: sentence-transformers/all-MiniLM-L6-v2
• Vector Store: Qdrant