Module 2 · Prompting, RAG & Agents revision

Revision guide · ~35 min · dense bullets, not full lessons

Chapters 2.1–2.9. Skim headings; drill weak spots in full lessons.

2.1–2.2 LLMs, decoding, APIs, structured outputs

• Pretrain (next-token on huge text) → finetune / align → infer (API call).
• Context window = hard budget for system + history + retrieved docs + output.
• Decoding knobs: temperature, top-p, max tokens, stop sequences. Determinism ≠ guaranteed without seed + temp 0 (still provider quirks).
• Structured outputs: JSON schema / tool schemas / constrained decoding. Always validate; never trust raw model text as typed data.
• Task prompts: summarize / QA / classify — specify format, audience, constraints, examples.

2.3–2.5 Evals, safety, advanced prompting, security

Evals

• Offline: gold sets, rubrics, LLM-as-judge (biased — calibrate). Online: thumbs, task success, latency, cost.
• Regression: freeze a prompt suite; fail the PR if quality drops.

Prompt patterns

• System role = durable policy. User = task instance. Few-shot = show pattern, not novels.
• CoT for multi-step reasoning; ask for final answer in a fixed format for grading.
• Contracts beat vibes: fields, enums, length, “if unknown say UNKNOWN”.

Security

• Prompt injection: untrusted content (web, PDF, email) steers the model. Separate trust boundaries; don’t let tools run on attacker text blindly.
• Least privilege for tools. Log tool calls. Human-in-the-loop for irreversible actions.

2.6–2.8 RAG fundamentals → production

Pipeline

query → embed → retrieve top-k → (rerank) → stuff into prompt → generate (+ cite)

Chunking

• Too big → noisy context. Too small → lose meaning. Overlap helps boundaries. Metadata (source, section) enables filters.

Indexes (know the tradeoffs)

• Flat / brute force — exact, slow at scale.
• IVF — cluster centroids; search nprobe lists. Faster, approximate; tune recall vs latency.
• PQ (product quantization) — compress vectors into codes; huge memory win, more approx error.
• HNSW — graph ANN; strong recall/latency; RAM heavy. Common default in vector DBs.

Production RAG

• Hybrid search (BM25 + dense). Rerankers when top-k is cheap noise.
• Citations & grounding: force quotes / IDs; refuse when retrieval weak.
• Eval: faithfulness, answer relevance, context precision/recall — not only “sounds good”.
• Ops: version embeddings + index; watch empty retrieval, p95 latency, cost/query.

2.9 Agentic AI & multi-agent

• Chatbot = reply in thread. Agent = plan + tools + memory + stop condition.
• Loop: observe → think/plan → act (tool) → observe → … → answer or escalate.
• Tools: typed functions; validate args; timeouts; idempotency for side effects.
• Memory: short-term (scratchpad / messages) vs long-term (store + retrieve). Don’t dump everything into context.
• Multi-agent: specialize roles; orchestrate (router, sequential, hierarchical). More agents ≠ more quality — more failure modes.
• Reflection / self-critique: second pass that checks against tools or rubric; cap retries.

30-minute drill

1. Draw IVF vs HNSW in one sentence each + one failure mode.
2. Write a 5-line system prompt that blocks tool use on untrusted pasted text.
3. List three RAG metrics you’d put on a dashboard before launch.