You can significantly improve Text-to-SQL accuracy without retraining models by systematically capturing and reusing error patterns as retrievable hints—Tahoe achieved 79% pass rate (up from 62%) on Spider 2.0 by learning from just 113 examples.
Tahoe is a system that improves Text-to-SQL performance by learning from errors and building a reusable library of hints. Instead of retraining models, it captures compiler feedback and user corrections as structured hints that guide LLMs to generate correct SQL for specific databases and dialects.