GoogleCloudPlatform/cymbal-air-toolbox-demo

What it solves

This project provides a production-quality reference implementation for building agentic applications that can query and interact with data stored in cloud databases. It specifically addresses the challenge of reducing hallucinations and providing agents with access to real-time, private, or specialized data that was not part of their original training set.

How it works

The system uses a three-tier architecture to enable an LLM to interact with databases securely:

1. Application: A user-facing app that orchestrates the interaction between the user and the AI agent.
2. MCP Toolbox: A middleware server that acts as an intermediary, exposing database operations as a set of discrete tools (e.g., find_flights). This layer handles authentication, authorization, and connection pooling, preventing the agent from accessing the database directly.
3. Database: A scalable cloud database containing the information the agent needs to answer queries.

The application employs Retrieval Augmented Generation (RAG) to ground responses in factual data and Agent-based Orchestration, where the LLM dynamically decides which tools to use and in what order to resolve a user's request.

Who it’s for

Developers looking for a blueprint on how to build secure, scalable agentic applications that integrate LLMs with cloud databases using a tool-based middleware approach.

Highlights

• Secure Database Access: Uses a middleware toolbox to prevent direct LLM-to-database connections.
• RAG Integration: Reduces hallucinations by augmenting prompts with retrieved database information.
• Agentic Reasoning: Moves beyond static chains of calls to allow the LLM to reason about tool selection.
• Interchangeable Database: Designed to be compatible with various preferred cloud databases.