Problem statement

Design the first production-ready slice of an internal chatbot: what users see, what you log, what happens when the model or DB fails.

Outcomes

• Streaming vs blocking tradeoffs; disclaimers; feedback affordances.
• Logging fields per turn (query, sources, model version, latency).

Deliverables

• chat_spec_onepager.md — stakeholders, P95 latency target (with rough justification), fields logged.

Links

• Phase 5

• [ ] Provider docs: streaming responses.
• [ ] Provider docs: rate limits and retries.

Time: ~2 h.

Task A — One-pager spec

chat_spec_onepager.md must include:

• Primary user + job-to-be-done
• Success metrics (quality + ops)
• P95 latency target (pick a number; justify order-of-magnitude)
• 10 logging fields minimum (timestamp, session id, question text hash, retrieved source ids, …)

Task B — Failure UX

Three bullets each: what UI shows for model timeout, empty retrieval, rate limit 429.

Submit: my-submissions/week-09/.

25 min.

1

Why is logging non-optional for shipped LLM features?

2

Name three fields you’d log for debugging a bad answer.

3

Streaming improves perceived latency—what risk does it add for compliance or citations?

Self-grade: Phase 5 checklist.

Concepts & curated links: read the Learn chapter Chatbots first (topic path, not calendar-only). This file is the weekly practice bank — extra drills and a checklist on top of Readings → Practice → Exam.

1. Practice bank

2. Week checklist

• [ ] Provider docs skim
• [ ] PRACTICE.md + EXAM.md