Yoomoney has developed a Text2SQL service that allows users to query analytical data using natural language. The system processes requests by identifying the intent, selecting relevant tables, generating SQL queries, and validating them before execution.
The service utilizes LangGraph to manage a non-linear pipeline of nodes. It features a two-level memory system that stores individual query snapshots and periodically updated summaries of the entire dialogue to maintain context.
To ensure accuracy, the system employs a four-layer validation subsystem. It checks for destructive operations, validates SQL syntax against MS SQL Server, performs semantic checks, and verifies the final output for logical consistency.
The platform includes a specialized tracing system to monitor silent failures. It records branching decisions, suppressed exceptions, data filtering events, and performance metrics, which are then logged to Langfuse and internal monitoring tools.
The team maintains a dedicated testing environment with a custom endpoint. This allows for experimental runs where system behaviors, such as validation layers or visualization modes, can be modified to evaluate performance.
