Documentación del TFG: Interconexión entre Espacios de Datos e I…

What is Documentación del TFG: Interconexión entre Espacios de Datos e Inteligencia Artificial Generativa?

This project documents a Final Degree Project (TFG) that designs and implements a multi-agent architecture interconnecting data spaces and generative AI. It uses a broker based on the Model Context Protocol (MCP) combined with the Google A2A (Agent to Agent) protocol, enabling structured communication between LLM agents and a local data space.

How to use Documentación del TFG: Interconexión entre Espacios de Datos e Inteligencia Artificial Generativa?

The system is deployed via Docker Compose. Agents register with the MCP broker on startup using a POST request to /agent/register. A command‑line client script (consola) sends natural‑language questions to the LLM agent, which generates SQL and coordinates the query cycle. Configuration is managed through a .env file with fixed agent IDs.

Key features of Documentación del TFG: Interconexión entre Espacios de Datos e Inteligencia Artificial Generativa

• MCP broker acts as a secure, modular hub for all agent communication.
• Agents communicate exclusively through the A2A protocol via the broker.
• Local LLM (TinyLlama) generates SQL from natural‑language questions.
• DuckDB serves as the local Iceberg‑formatted data space.
• Containerized architecture with FastAPI REST interfaces.
• CLI client for terminal interaction.

Use cases of Documentación del TFG: Interconexión entre Espacios de Datos e Inteligencia Artificial Generativa

• Querying sales data (fecha, producto, cantidad, precio) using natural language.
• Demonstrating agent‑to‑agent communication without direct agent‑agent or agent‑database connections.
• Testing an MCP‑based broker for interoperable multi‑agent systems.
• Prototyping a local, reproducible data‑space architecture with generative AI.

FAQ from Documentación del TFG: Interconexión entre Espacios de Datos e Inteligencia Artificial Generativa

What distinguishes this architecture from direct LLM–database interaction?

All data access must pass through the MCP broker as a “tool”, enforcing a strict separation between semantic processing (LLM) and data retrieval (agent).

What runtime dependencies are required?

The system uses Docker Compose, DuckDB, FastAPI, and Transformer’s TinyLlama-1.1B-Chat-v1.0. No external cloud dependencies are needed.

Where does the data reside?

Data is stored locally in a DuckDB file (lake.duckdb) containing the table iceberg_space.ventas. It is loaded via the load_data.py script.

What are the known limitations?

The current implementation is a local prototype. Future work includes adopting full Google A2A specifications (Agent Card, JSON‑RPC 2.0, SSE), persisting messages and agents to a database, and adding conversation IDs for traceability.

How is agent identity managed?

Agents register at startup with a fixed agent_id from the .env file (or an auto‑generated one). The broker maintains a registry in memory.