AI 产品与体验

Great AI UX reduces confusion and builds trust. This chapter covers product patterns for LLM features.

1) Input UX

• Guide rails: suggest prompts/templates; inline tips for constraints (length, format).
• Clarify capability limits and data sources.
• File input: show size/type limits; progress + error states.

2) Output UX

• Streaming for responsiveness; show typing indicator.
• Source citations: link to documents/lines; highlight referenced snippets.
• Expandable details: show reasoning/steps optionally; keep default concise.
• Offer quick actions: copy, save, refine, thumbs up/down.

3) Correction & Refinement

• One-click refine prompts (shorter/longer, tone change, translate).
• Ask for missing info when context insufficient; don’t hallucinate.
• For structured tasks, allow editing of fields then re-run.

4) Safety & Expectations

• Disclaimers about limitations; show model name/version.
• Guardrails: refuse unsafe requests with clear messaging.
• Privacy notice: what’s sent to the model; link to policy.

5) Error States & Recovery

• Friendly errors with next-step suggestions.
• Retry/alternate model option when provider fails.
• Partial results: show what succeeded, not just a failure screen.

6) Performance Cues

• Show token/usage summary when relevant; warn on large inputs.
• Latency feedback: skeleton/loading states; keep UI interactive.
• Offline/slow mode: queue requests or local fallback if available.

7) Personalization & Memory

• Remember user preferences (language/tone) per session/tenant with consent.
• Allow opt-out of history retention; easy clear/reset.
• For shared contexts, indicate whose data is used.

8) Metrics

• Task success rate, edit/refine rate, abandonment, feedback scores.
• Track which prompts/templates are used; prune low performers.
• A/B outputs formatting and copy to reduce confusion.

9) Minimal Checklist

• Streaming + citations + refine actions.
• Clear limits, errors, and safety messaging.
• Feedback hooks + metrics on success/abandonment/refine.