mcp_poc

An implementation of a RAG (Retrieval-Augmented Generation) system utilizing the Model Context Protocol (MCP). This project demonstrates a modular architecture for legal document processing and retrieval using ChromaDB as a vector store and Ollama for local LLM inference.

🏗 Project Structure

The project is organized into several microservices:

• mcp_server/: The core MCP server that exposes semantic search tools over the vector database. This is the primary service — connect any MCP-compatible LLM client directly to it.

• ingestion/: REST API service for fetching and embedding documents from the SAOS (court judgments) and ELI (legislative acts) APIs into ChromaDB.

• frontend/: Optional local chat UI backed by Ollama and the MCP server.

• data/: Local storage for the ChromaDB database and other persistent assets.

• scripts/: Utility scripts for ingestion and maintenance.

Related MCP server: Legal MCP

🚀 Getting Started

Prerequisites

• Docker & Docker Compose: Required for containerized deployment.

• Python 3.10+: For local development.

• Ollama: Installed locally, or use the integrated service in docker-compose.yml.

• NVIDIA Container Toolkit: (Optional) For GPU acceleration within Docker.

Environment Setup

1. Clone the repository:

   git clone https://github.com/barwojcik/legal_mcp.git
   cd legal_mcp

2. Copy and review the environment variables:

   cp .env.example .env
   # Edit .env if you want to use OpenAI/Google embeddings instead of Ollama

Running with Docker Compose

Spin up ChromaDB, Ollama, the MCP server, and the ingestion service:

docker compose up -d chroma ollama mcp-server ingestion

To also run the optional frontend:

docker compose up -d

Services will be available at:

• ChromaDB: http://localhost:8000

• Ollama: http://localhost:11434

• MCP Server: http://localhost:8001/mcp

• Ingestion API: http://localhost:8002

• Frontend (optional): http://localhost:8003

Populate the database

bash scripts/ingest_saso.sh

This fetches one page (20 judgments) from the SAOS API and embeds them into ChromaDB. See scripts/ingest_saso.sh and the Ingestion API docs for more options.

🔌 Connecting a Commercial LLM

The MCP server speaks the Model Context Protocol over HTTP/SSE. Once the stack is running, point your LLM client at http://localhost:8001/mcp.

Claude Desktop

Add the following to your claude_desktop_config.json (usually at ~/Library/Application Support/Claude/claude_desktop_config.json on macOS or %APPDATA%\Claude\claude_desktop_config.json on Windows):

{
  "mcpServers": {
    "legal": {
      "url": "http://localhost:8001/mcp",
      "type": "http"
    }
  }
}

Cursor / Zed / other MCP clients

Add an MCP server entry pointing to http://localhost:8001/mcp. Refer to your client's documentation for the exact configuration format.

Once connected, the LLM will have access to 14 tools for searching and retrieving Polish court judgments and legislative acts.

🛠 Ingestion API

The ingestion service exposes a REST API at http://localhost:8002.

Ingest SAOS court judgments:

curl -X POST http://localhost:8002/update \
  -H "Content-Type: application/json" \
  -d '{"n_pages": 1, "page_size": 20}'

Ingest ELI legislative acts:

curl -X POST http://localhost:8002/eli-update \
  -H "Content-Type: application/json" \
  -d '{"n_pages": 1, "page_size": 20}'

🛠 Development

Linting and Type Checking

# Run ruff
ruff check . --fix

# Run mypy
mypy .

Pre-commit Hooks

pre-commit install

⚖ License

This project is licensed under the Apache-2.0 licence.