Week 8. RAG failure modes + timed midterm checkpoint.
Problem statement
RAG went to staging. Legal asks about wrong citations; Support asks about stale docs. You document failure modes and mitigations—and complete the midterm checkpoint.
Outcomes
- Hybrid search / reranking / ACL-aware retrieval — conceptual.
- Two observed failure cases from Week 07 traces with mitigations.
Deliverables
rag_failure_modes.md— two failures + detection/mitigation.MIDTERM.md— timed exercise (see file).
Links
- Graduate midterm (parallel): TPM-PM-AI-Graduate-Course.md (see section Midterm)
- Phase 4
- [ ] Hybrid search (BM25 + dense) — one blog or doc chapter.
- [ ] Indirect prompt injection via retrieved content — one OWASP or vendor note.
Time: ~2 h.
Task A — Failure analysis
From your Week 07 rag_trace.md, pick two weak answers. For each: root cause (chunking, stale, wrong retrieval), how you’d detect in prod, mitigation.
Save rag_failure_modes.md.
Task B — Midterm
Open MIDTERM.md and complete under stated time limit.
Submit: my-submissions/week-08/.
If you already completed MIDTERM.md, treat that as your Week 08 exam. Otherwise:
Alt exam (20 min)
- Define indirect prompt injection with a RAG example.
- Why is citation fidelity a product requirement, not only accuracy?
- Name one benefit of hybrid search.
Self-grade: Phase 4 checklist.
Total: 90 minutes. Stop when time is up. Use only: SQLite + practice-env/data/cloudnote.db, schema from practice-env README, and blank paper/notes you wrote yourself.
Part A — Metric precision (30 min)
Rewrite these ambiguous questions into testable metric definitions (numerator, denominator, window, segment if needed):
- “Did the AI sidebar hurt retention?”
- “Are enterprise customers unhappy?”
- “Is the product getting slower?”
Part B — SQL (45 min)
On cloudnote.db, write queries that:
- Count distinct users with at least one
checkout_completeevent, by segment. - List top 3
event_namevalues by count in February 2025 (useevent_tstext comparison>= '2025-02-01' AND < '2025-03-01'or equivalent).
Run queries and paste row counts or sample rows.
Part C — Short concepts (15 min)
Define in ≤ 2 sentences each: embedding, RAG, prompt injection, temperature.
Afterward: Grade using answer key you write tomorrow with references open; update weak topics.
Concepts & curated links: read the Learn chapter RAG 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
| # | Problem | Data / materials | Expected outcome |
|---|---|---|---|
| P1 | Document 2 weak RAG answers from Week 7; root cause + mitigation. | Your rag_trace.md |
rag_failure_modes.md |
| P2 | Complete MIDTERM.md under time limit. |
cloudnote.db |
Self-grade next day |
2. Week checklist
- [ ] Failure analysis + midterm
- [ ]
PRACTICE.md+EXAM.md