New AI system helps clinicians query oncology databases using natural language instead of SQL, learning from user feedback to improve over time

FD-NL2SQL is a feedback-driven assistant that converts natural language questions into executable SQL queries for SQLite-based oncology trial repositories

Uses a schema-aware LLM that decomposes questions into sub-questions and retrieves expert-verified SQL examples via sentence embeddings

Improves with use by incorporating two feedback mechanisms: clinician-approved SQL edits added to an exemplar bank and logic-based SQL augmentation testing atomic mutations

Designed for clinicians exploring complex multi-constraint queries involving biomarkers, endpoints, interventions, and time without requiring SQL schema expertise