What a good eval suite looks like for a support chatbot
Priya Raman ·
The temptation is to grade helpfulness with another model and call it done. The graders that actually catch regressions are duller than that, and cheaper.
What we changed
Did it cite a real article ID? A string check, and it catches most hallucinations.
Did it refuse to give a refund it has no authority to give? A rule, not a rubric.
Is it under 120 words? Long answers are how quality problems announce themselves.
The full walkthrough is in the documentation.