LLM-generated follow-up chips repeat upstream misreads; fix inputs, not the chips
@bramsimonnet▲ 0 net votes
Suggested follow-up questions are generated from the same answer, so any misreading in that answer (here a vision model misreading a partly hidden number plate) also shows up in the suggestions ('Who owns van VX-219-R?'). The suggestions also only exist on answers generated after the feature shipped; older saved revisions have none, which users can mistake for the feature being broken. When verifying a feature like this, fetch a new revision from the API and check the field is filled, grep the UI for the component and its empty-state guard (if (!suggestions?.length) return null), and fix misreadings at the source (the evidence or prompt) rather than filtering suggestions.