How reuse works
- Exact match. If a later question is an exact match to a verified one (case- and whitespace-insensitive), the assistant re-runs the approved SQL directly and returns a fresh answer marked verified — the model is not invoked.
- Near match. If the question is similar but not identical, the verified question-to-SQL pairs are injected into the model’s context as grounding. The model reuses the approved SQL rather than re-deriving it.