Capture user feedback and link it to the logged call

Offline evals miss what real users do. A lightweight feedback signal closes that gap.

Pick a signal that costs the user almost nothing:

• thumbs up or down
• an edit to the generated text
• whether the user accepted a suggestion

Store it with the id of the logged call, so every piece of feedback points to the exact prompt, model and input behind it.

Review negative feedback weekly. Each pattern you find becomes new eval cases. Over time, the share of positive feedback tells you whether changes help in practice, not just on your test set.