Problem statement

Ship a thin demo: user asks a question; system uses corpus and/or read-only SQL; logs a turn. Write the incident playbook for launch night.

Outcomes

• Minimal Streamlit/Gradio/CLI acceptable.
• Playbook for wrong answer, DB down, rate limit.

Deliverables

• Runnable artifact + README_run.md.
• incident_playbook.md (≤ 1 page) covering model wrong, DB down, rate limit.

Links

• Phase 5 mini deliverable

• [ ] Streamlit or Gradio quickstart (pick one).
• [ ] Skim one doc on prompt injection in chat UIs.

Time: ~2–3 h.

Task A — Minimal chat path

Implement one of:

• Streamlit/Gradio app that accepts a question and shows answer + sources (RAG from practice-env/corpus/ or canned retrieval).
• CLI Python script with same contract.

Log each turn to a file or stdout with latency.

Task B — Incident playbook

incident_playbook.md: for model wrong, DB unavailable, 429 rate limit — detection, first response, escalation, comms template (one line each).

Task C — Demo

Record a 2-minute screen capture or paste a transcript of one happy path + one failure path.

Submit: my-submissions/week-10/ (code + docs).

30 min.

1

When is naive RAG chat enough vs when do you need an agent with tools?

2

Write 5 bullet runbook steps if users report “the bot invented revenue numbers.”

3

What is one PII risk in logging full user questions?

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

• [ ] UI framework hello-world
• [ ] PRACTICE.md + EXAM.md