Proyecto Tfg

🚀 Interoperability between Data Spaces and Generative AI

This project aims to design and implement a robust and scalable architecture for the interaction between language models and data spaces, ensuring secure and modular data access.

🚀 Quick Start

This README provides an in - depth overview of the project focused on designing and implementing interoperability between data spaces and generative AI. It details the project's purpose, current architecture, fundamental principles, future scalability, and conclusions.

✨ Features

🎯 Project Purpose

The main goal of this graduation project is to design and implement a powerful and scalable architecture that enables Large Language Models (LLMs) to interact with data spaces through the Model Context Protocol (MCP). The client (i.e., the LLM) should not directly access the database. Instead, all operations must be securely, modularly, and scalably performed through the MCP server.

🧱 Current Architecture

Main Components:

• 🧠 Client LLM (cliente_llm.py)

  • Execute natural - language questions.
  • Use the TinyLlama - 1.1B - Chat - v1.0 model.
  • Convert questions into SQL queries.
  • Query data through the MCP server.
  • Interpret the returned results and construct explanatory answers.
  • Log the entire process in the log file: logs/cliente_llm.log.

• 🔗 MCP Server (main.py)

  • Developed using FastAPI.
  • Connect to the local data space based on DuckDB.
  • Expose MCP tools through REST endpoints:
    • /tool/consulta: Execute SQL queries.
    • /tool/info/productos: Return a list of different products.
    • /tool/info/fechas: Return the registered minimum and maximum date ranges.

• 📂 Data Space

  • Implemented locally using DuckDB (lake.duckdb).
  • Contains the iceberg_space.ventas table with the following columns:
    • fecha (DATE)
    • producto (TEXT)
    • cantidad (INTEGER)
    • precio (DOUBLE)
  • Data is loaded from load_data.py.

🔒 Fundamental Principles and Key Decisions

• ✅ Strictly separate semantic processing (LLM) and data access (MCP).
• ✅ Comply with the MCP design: LLMs can only access data through "tools".
• ✅ Use rich context prompts containing prior information obtained from the MCP.
• ✅ Adopt a modular, scalable, and log - traceable architecture.

📈 Future Scalability

The current design takes future development into account from the very beginning:

• 🔁 Replace DuckDB with Apache Iceberg or Trino/Presto in the future.
• 🤖 Replace TinyLlama with a more advanced LLM or a GPU - deployed model in the future.
• 🔎 Evolve into a more advanced Retrieval - Augmented Generation (RAG) architecture, where:
  • The LLM first retrieves content through a vector store based on embeddings generated from the MCP.
  • The retrieved content is passed to the model as context to obtain more accurate answers.

In addition, new MCP tools can be integrated, such as:

• /tool/info/esquema
• /tool/info/documentacion
• /tool/descargar
• /tool/upload - pdf

📜 Conclusion

A solid, powerful, and compliant foundation has been established. The system has achieved complete interaction between the language model and the data space, complying with MCP principles and paving the way for future development towards a more advanced RAG architecture.

📚 Documentation

Project Information