BAP-SQL Improves Text-to-SQL Agent Efficiency with Budget-Aware Planning.

Large language model agents that interact with databases often struggle with inefficient query execution, consuming excessive context and database resources before finding relevant information. Traditional methods of compressing data after it's retrieved cannot fix these initial inefficiencies. This research proposes BAP-SQL, a novel approach that treats the process of forming observations as a critical budget-management stage. BAP-SQL works by first assessing the potential risk of a database query and then intelligently rewriting SQL queries to be more efficient. It also incorporates a separate runtime component that enforces strict budget limits. The system demonstrated improved success rates, particularly when operating under tight resource constraints, and achieved these gains with fewer tokens compared to existing methods.