Building a Multi-Agent Pipeline for NL-to-SQL Analysis

The existing MCP (Model Context Protocol)-based NL-to-SQL system had a one-directional structure in which SQL generation and execution were separated, and the entire process was handled by a single engine in a monolithic manner. This made it difficult to identify the cause when errors occurred, and as rules increased, the prompt became bloated, making maintenance increasingly difficult.

To solve this, instead of concentrating all responsibility in one place, we designed an A2A (Agent-to-Agent) pipeline in which specialized agents collaborate step by step. In this structure, natural language interpretation, query generation, execution, and result explanation are each carried out by independent agents.

Through the A2A architecture, we secured the following benefits:

• Each stage is isolated, enabling immediate identification of the cause when an error occurs
• Even when new rules or features are added, the impact on existing behavior is minimized
• An analytical foundation with accuracy, consistency, and operational stability is established

For analysis, we use raw event logs from UTS (User Tracking System), which richly capture the context of user behavior. UTS provides detailed user behavior data for various services through a JSON-based denormalized structure, and since it shares the same schema rules, it is easy to expand the scope of analysis.