Quick reference · AI evaluations
Design an eval that answers a real question
An evaluation (eval) checks how an AI system behaves on examples chosen to represent a real task.
- Name the task. What should the system do for a user?
- Choose cases. Include ordinary requests, edge cases, and cases where the right behavior is to refuse or ask a question.
- Write pass criteria. Describe observable properties of a good result, not a vague goal like “be smart.”
- Choose a grader. Use exact checks for exact requirements, a rubric for judgment, or human review when errors matter or graders are uncertain.
- Inspect failures. Read examples behind the scores. Group errors by cause before changing the system.
- Repeat. Rerun the same cases after changes, then add new cases for failures you discover.
Keep different questions separate
| Question | Example measurement | What it cannot prove alone |
|---|---|---|
| Did retrieval find the needed source? | Recall at k | That the final answer is correct |
| Did the answer follow the evidence? | Grounding or citation checks | That the evidence itself is true or complete |
| Did it answer the user’s actual question? | Task-specific rubric or human review | That it will work on every future request |
| Did the service work reliably? | Failure rate and latency | Answer quality |