04. Week 7 — Daily Recall¶
Spaced practice. Answer from memory. If you get stuck, jump to the explainer section in parentheses.
Monday (after ELI5 + Chapter 1)¶
- In the library analogy, who are the librarian, bookshelf, index card, reading desk, and answer brief? (ELI5)
- Why is a confident wrong answer about Q4 revenue more dangerous than a polite refusal? (§1.1)
- Why do closed-book LLMs hallucinate on company-specific facts? (§1.2)
- RAG in one sentence: what changes at answer time? (§1.4)
- Explain the difference between a model sounding sure and a system being grounded. (§1.5)
Tuesday (after Chapter 2)¶
- Why can’t you just stuff the whole PDF into the prompt every time? (§2.1)
- Chunk size trade-off: what breaks when chunks are too small? Too large? (§2.2)
- What does overlap protect against? Give one concrete example. (§2.3)
- Recursive splitting vs semantic splitting — what is the basic difference? (§2.4)
- Same document chunked three ways: which version would best answer a precise refund-policy query? Why? (§2.5)
- What metadata would you store beside each chunk? (§2.6)
Wednesday (after Chapter 3)¶
- What does an embedding represent in plain language? (§3.1)
- Why do semantically similar sentences land close in embedding space? (§3.2)
- Cosine similarity vs dot product — when do they behave the same? (§3.3)
- Name four things you would check before choosing an embedding model. (§3.4)
- HNSW in one sentence: what structure does it use? (§3.5)
- IVF in one sentence: what idea makes it fast? (§3.5)
- Why do numbers, negation, and acronyms often trip embeddings? (§3.6)
Thursday (after Chapter 4)¶
- Draw the naive RAG pipeline from memory. (§4.1)
- Give one failure mode for query understanding, retrieval, reranking, and prompt augmentation. (§4.2-§4.6)
- Why is reranking useful even after you already have top-k retrieved chunks? (§4.5)
- What should the augmented prompt force the model to do when evidence is missing? (§4.6)
- Write one retrieval prompt that asks the model to rewrite a user query for search. (§4.8)
- Why does RAG not automatically solve multi-hop reasoning? (§4.9)
Friday (after Chapter 5)¶
- Recall@k — what exactly are you counting? (§5.2)
- MRR — what does it reward that recall@k ignores? (§5.3)
- NDCG — why do we discount lower-ranked items? (§5.4)
- Faithfulness vs answer relevance — explain the difference. (§5.5)
- What does RAGAS measure well, and why is human review still needed? (§5.6)
- Why do teams get fooled when they evaluate only answer fluency? (§5.1)
Weekend (pre-revision)¶
- Sketch the failure-fix table from explainer §6.1 from memory.
- Give a production latency budget for embed → retrieve → rerank → generate. (§6.4)
- Answer five interview questions from explainer §6.3 without notes.
- Which five basics must feel automatic before
09_advanced_rag_patterns? (§6.6) - Say the bridge sentence to the next module out loud. (§6.7)