Problem statement

You join an exec briefing where engineers say: “We’re hitting context limits; inference cost scales with tokens.” You must recognize every term, ask one clarifying question each about cost and reliability, and explain why “dump the whole wiki into the prompt” cannot work.

Outcomes

By Friday you can define tokens, context window, inference vs training, hallucination vs grounding, embeddings, and fine-tuning vs RAG without reading slides.

Deliverables

1. READINGS.md completed (checkboxes).
2. PRACTICE.md exercises + artifact in my-submissions/week-01/ (or this folder).
3. EXAM.md closed-book attempt before opening notes.

Files in this folder

• READINGS.md — provider docs + OWASP skim.
• PRACTICE.md — glossary, context-limit exercise using data/.
• EXAM.md — short oral/written check.
• data/long_briefing_stub.txt — optional prop for “why truncation matters.”

Links

• Master plan: Phase 0
• Shared DB setup (needed this week): practice-env

Complete before PRACTICE.md. Spend ~2–3 hours total.

Required (pick one provider ecosystem)

Skim official docs for:

• [ ] How tokens are counted (rough rule: input + output).
• [ ] Chat completions vs legacy naming (whatever your provider calls multi-turn chat).
• [ ] JSON / structured output mode (schema or constrained JSON).
• [ ] Any page on models listing context window sizes for the models you might use.

Suggested entry points (swap for your stack):

• OpenAI: platform docs on tokens, chat, structured outputs.
• Anthropic: Messages API, model cards / context limits.
• Google AI Studio / Gemini: context length, JSON mode.

Recommended

• [ ] OWASP LLM Top 10 — read the titles + one paragraph each, not memorization.

Optional

• [ ] One blog post or doc on embedding vs generation (why retrieval uses embeddings).

Reading notes: After each session, write 5 bullets from memory (no paste). Store in your submissions folder.

Lab A — Environment (required)

Follow practice-env/README.md:

1. Create venv, pip install -r requirements.txt.
2. Run python scripts/build_dummy_db.py.
3. Run python labs/explore_metrics.py.
4. Save terminal output (or screenshot) + one sentence: what the rollup measures.

Lab B — Glossary (required)

Create glossary.md with 10 terms from the master plan’s jargon list (Phase 0 cheat sheet). Define each in your own words in ≤ 2 lines.

Lab C — Context limits (required)

1. Open data/long_briefing_stub.txt. Imagine you must fit “the key decisions” into ≤ 150 words for a VP.
2. Write that summary yourself (no LLM), then optionally compare with one LLM summary.
3. Write 5 bullets: what you would drop first if the context budget shrinks again—tie drops to business risk (e.g. losing audit trail vs losing narrative).

Lab D — Short essay (required)

In ≤ 250 words, explain why context limits force summarization, chunking, or RAG for large corpora. Link to a product consequence (cost, wrong answers, stale policy).

Submit: folder my-submissions/week-01/ containing glossary.md, essay_context_limits.md, env proof, optional comparison notes.

Rules: Answer without the learning plan or web open (30–45 minutes). Afterward, compare with Phase 0 checklist and restudy gaps.

Section A — Definitions (20 min)

Define in 2–3 sentences each:

1. Token
2. Context window
3. Inference vs training
4. Hallucination vs grounding
5. Embedding (what it’s for)
6. Fine-tuning vs RAG (when each is a reasonable bet)

Section B — Scenario (15 min)

An engineer proposes: “We’ll paste all 500 pages of our internal wiki into the prompt so the model knows everything.” Write 4 sentences explaining why this fails and two mitigations (pick from summarization, chunking, RAG, structured retrieval).

Section C — Vocabulary in the wild (10 min)

In a mock Slack message, use correctly in one short paragraph: guardrails, prompt injection, eval, latency. (Invent context if needed.)

Self-grade: Meets = all definitions accurate; scenario names tradeoffs; Slack paragraph uses terms correctly. Restudy anything marked “partial.”

Concepts & curated links: read the Learn chapter Foundations 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

Use practice-env for anything marked with data below. Paths are relative to the practice-env/ folder in this repo.

Expanded (same as table)

P4 — Environment

• Problem: Prove the practice database runs end-to-end.
• Data: SQLite cloudnote.db, script output.
• Outcome: Reproducible proof you can load metrics locally (foundation for Weeks 4–6).

P5 — Context / summarization

• Problem: Practice aggressive summarization tradeoffs.
• Data: Long stub briefing.
• Outcome: Explicit list of what you dropped and why (ties to context limits).

2. Week checklist

• [ ] Watched/read at least one resource per topic row (T1–T6)
• [ ] Finished P4 (practice env)
• [ ] Finished P5 or P1–P3 written prompts
• [ ] Completed READINGS.md and PRACTICE.md for the formal week assignment
• [ ] EXAM.md without peeking at notes first