Before you start: Familiarity with retrieval.
Build the set
- Collect real questions.
- Include questions that must be refused or handed off.
- Write the expected facts, not exact wording.
Grade
- Check required facts appear.
- Check forbidden claims do not.
- Track pass rate over time and on every change.
Use it
Run before releases and on a schedule. A drop tells you something changed in the data, prompt or provider.
Part of: Developer fundamentals
Next recommendedHow retrieval grounds an answerRelated
Last updated 2026-09-25. Something unclear or wrong? Tell us.