PRODUCTION SCENARIO
A Generative AI Engineer's assistant scores 0.95 on Correctness against the team's evaluation dataset before release. In its first week in production, support logs 40 wrong answers, none of which resemble anything in that dataset.
What should the engineer do?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Add the failing production traces to the evaluation dataset and re-score
The offline dataset never contained the inputs that fail, so no threshold or judge applied to it can predict production. Evaluation dataset records can be created from production traces and keep a link back to the source trace, so the fix is to fold the failures in and re-score. Raising the gate tightens a measure of unrepresentative data, a Guidelines judge changes the quality dimension rather than the sample, and a higher sample rate scores more live traffic without adding the expectations a correctness check needs.