Text-to-SQL in Production: Schema Linking, Value Retrieval, and Execution-Guided Self-Correction
Translating natural language into executable database queries is one of the most widely deployed applications of large language models in enterprise software. It is also one of the most brittle. On synthetic academic benchmarks such as Spider 1.0, frontier models regularly exceed 90% execution accuracy. However, evaluating those same models on realistic enterprise estates yields a steep drop. On the Spider 2.0 benchmark, which evaluates real-world data warehouses spanning BigQuery and Snowflake
1 min