Error analysis

Error analysis is the step before writing evals. You pull a sample of real inputs and outputs, read them one by one, and jot a short note on every problem you see. Then you group the notes into failure categories and count them. The result is a ranked list of what's actually going wrong, which is rarely what you'd have guessed.

It matters because generic quality metrics miss product-specific failures. You might find the summary feature keeps dropping the customer's deadline, that extraction breaks on forwarded emails, or that most bad answers trace back to retrieval rather than the prompt. Each category points to a fix and to a check you can automate: a code assertion, a grader prompt, or new golden-dataset cases.

The mistake is jumping straight to dashboards and model graders without reading outputs. Keep reading until new categories stop appearing, and involve someone who knows the domain, because they spot errors others miss. Full traces make the work much easier, since you can see the retrieved context and intermediate steps, not just the final answer. Repeat after every significant change, and whenever user feedback shifts.