eu-ai-act-assistant

A Retrieval-Augmented Generation (RAG) agent designed to answer questions about the EU AI Act using a local vector index and specific tools. The project consists of a frontend interface (Gradio) and a backend component providing tools via the Model Context Protocol (MCP).

Project Structure

The project is split into two main components:

• Frontend (/frontend):

  • Interface: Built with Gradio (gradio_agent.py).
  • Functionality:
    • Interacts with the backend MCP server.
    • Uses LangChain for agentic logic.
    • Intelligent Memory: Integrated with Supermemory for persistent and semantic conversation memory.
    • RAG Evaluator: Integrated evaluator (rag_evaluator.py) that scores answers for Faithfulness and Answer Relevancy using ragas.
  • Configuration: Uses Dockerfile to build the agent image.

• Backend (/backend):

  • Server: An MCP server (rag_mcp_server.py) exposing retrieval tools.
  • Index: Contains the pre-built vector index (eu_ai_act_index) and the indexer script (eu_ai_act_indexer.py).
  • Configuration: Uses Dockerfile to build the MCP server image.

Prerequisites

• Docker and Docker Compose installed.
• An OpenAI API Key.
• A Supermemory API Key (optional but recommended for persistent memory).

Getting Started

1. Set Environment Variables Ensure your API keys are available in your shell environment, or set them in the docker-compose.yml file.

   export OPENAI_API_KEY=your_openai_key_here
   export SUPERMEMORY_API_KEY=your_supermemory_key_here

   [!TIP] You can obtain a Supermemory API key from console.supermemory.ai.

2. Run with Docker Compose Navigate to this directory and run:

   docker-compose up --build

   This command will:

   • Build the mcp-server image from the backend directory.
   • Build the rag-eu-agent image from the frontend directory.
   • Start both containers.

3. Access the Application Open your browser and navigate to:

   http://localhost:7860

4. Usage

   • Type your question about the EU AI Act in the chat box.
   • The agent will use the eu_ai_act_retriever tool to find relevant context.
   • Memory: The agent retrieves relevant context from past conversations via Supermemory to provide personalized answers.
   • Evaluation: Shortly after the answer is complete, the "Evaluation Metrics" box below the chat will update with scores for the response's faithfulness and relevancy.

Memory Management (Supermemory)

The assistant uses Supermemory to move beyond simple chat history. Instead of just remembering the last few messages, it:

• Indexes Conversations: Every interaction is stored and semantically indexed.
• Semantic Retrieval: Relevant context from any previous chat is retrieved based on your current question.

To disable memory, simply omit the SUPERMEMORY_API_KEY.

Development

• Frontend Code: Located in frontend/gradio_agent.py.
• Evaluator: frontend/rag_evaluator.py.
• Backend Tools: Defined in backend/rag_knowledge_source.py.

Notes

• The setup uses a bridge network so the frontend can securely communicate with the backend MCP server.
• The vector index is mounted as a volume to the backend container to ensure data persistence and accessibility.