Evaluate quality beyond one successful example
Build test cases with inputs and expected results or evaluation criteria. Keep examples used during development separate from a reserved set for testing changes. Include normal cases, edge cases, missing information and adversarial inputs.
In classification, precision and recall measure different things; false positives and false negatives can have different costs. In retrieval, metrics such as recall@k assess whether relevant information was found. For generation, combine source checks, rules and human review. An AI-based evaluator also needs validation.
A new version may improve the average while making urgent requests worse. Compare results by scenario, review human disagreements and preserve the cases that exposed regressions.
Useful evaluation reflects the real decisions and risks of use.
Try it yourself
Design ten evaluation cases for a document assistant: include an answer that is present, one that is absent, conflicting sources, an outdated document and a document the user is not authorised to read. Define how to distinguish each failure type.