# Chapter 2 Interview Questions ## Concept Foundations ### 1. What is the RAG query lifecycle? Strong answer: It is the journey from user question to retrieved evidence, grounded prompt, answer, citations, and evaluation. Red flag: "It is just retrieval." Follow-up: Which stage would you inspect first when an answer is wrong? What this proves: You see RAG as a pipeline. ### 2. Why should retrieval be evaluated separately from generation? Strong answer: The model cannot use evidence that retrieval fails to provide. Red flag: "Only the final answer matters." Follow-up: What retrieval metric would you start with? What this proves: You can debug stages independently. ### 3. What is query understanding? Strong answer: Interpreting intent, missing details, risk, and filters before searching. Red flag: "Use the exact user text every time." Follow-up: When should the assistant ask a clarification question? What this proves: You understand the request before search. ### 4. Why do metadata filters matter? Strong answer: They prevent semantically similar but wrong documents from being retrieved, such as wrong region or policy version. Red flag: "Embeddings handle everything." Follow-up: Give one metadata filter for a refund bot. What this proves: You understand governed retrieval. ## Production Judgment ### 5. What should be logged in a RAG trace? Strong answer: Request id, question, filters, retrieved source ids, scores, prompt version, answer status, citations, refusal reason, latency, and cost. Red flag: "Only log the final answer." Follow-up: Which logs are sensitive? What this proves: You think about observability. ### 6. When should a lifecycle include access checks? Strong answer: Before or during retrieval, and again before showing citations if sensitive sources are involved. Red flag: "Retrieve first, hide later." Follow-up: What can go wrong with post-generation-only access control? What this proves: You reason about security boundaries. ### 7. How do latency and cost enter the lifecycle? Strong answer: Query rewrite, retrieval, reranking, context size, and generation all add latency and cost. Red flag: "RAG cost is only the model call." Follow-up: Which step would you optimize first? What this proves: You understand production trade-offs. ### 8. What is context assembly? Strong answer: Selecting, ordering, trimming, and formatting retrieved evidence before sending it to the model. Red flag: "Paste all chunks into the prompt." Follow-up: How can bad ordering hurt the answer? What this proves: You know context is designed. ### 9. What is a helpful refusal in the lifecycle? Strong answer: A response that explains missing evidence and gives the next safe step, rather than guessing. Red flag: "Refuse with no explanation." Follow-up: Write one refusal for a missing policy source. What this proves: You see refusal as product UX. ### 10. How does reranking fit into the lifecycle? Strong answer: It reorders retrieved candidates so the strongest evidence enters the context window. Red flag: "Reranking is the same as embedding." Follow-up: What failure can reranking introduce? What this proves: You understand retrieval strategies. ## Evaluation and Failure Modes ### 11. What is the first debug step for a wrong answer? Strong answer: Inspect retrieved context and source ids before changing prompts. Red flag: "Tune the prompt first." Follow-up: What if the right source was retrieved? What this proves: You debug in order. ### 12. How do you catch stale index issues? Strong answer: Track document version, ingestion timestamp, freshness checks, and evals that expect the latest approved policy. Red flag: "Users will report it." Follow-up: What alert would you create? What this proves: You think operationally. ### 13. What does citation precision measure? Strong answer: Whether cited sources actually support the answer. Red flag: "Any citation is good." Follow-up: How can a citation be misleading? What this proves: You care about trust. ### 14. Why are multi-hop questions harder? Strong answer: They require evidence from multiple sources, and missing one source can produce an incomplete answer. Red flag: "Top one chunk is enough." Follow-up: How would you evaluate multi-hop retrieval? What this proves: You understand complex enterprise queries. ### 15. What is an execution pipeline? Strong answer: The ordered runtime path of checks and actions: auth, cache, query rewrite, retrieval, rerank, prompt, safety, answer, logging. Red flag: "It is just a script." Follow-up: Which step can be asynchronous? What this proves: You can design a shippable system. ## Architecture ### 16. Where do indexing strategies fit? Strong answer: Before query time, during ingestion and storage, but they shape retrieval quality during the lifecycle. Red flag: "Indexing is unrelated to querying." Follow-up: How does chunking affect indexing? What this proves: You connect offline and online systems. ### 17. Where do retrieval strategies fit? Strong answer: In the retrieval stage: vector, keyword, hybrid, filters, reranking, and query routing. Red flag: "There is only vector search." Follow-up: When is hybrid search better? What this proves: You understand retrieval options. ### 18. How would you explain the lifecycle to a PM? Strong answer: The system gathers approved evidence before answering, cites its sources, and refuses when it cannot prove the answer. Red flag: "It is embeddings plus LLM." Follow-up: What product metric matters? What this proves: You can communicate across roles. ### 19. What makes a lifecycle production-ready? Strong answer: Access control, versioning, evals, logging, latency/cost budgets, fallbacks, and clear ownership. Red flag: "It works in a notebook." Follow-up: Which item would you add first? What this proves: You can move from demo to product. ### 20. What should a Chapter 2 project prove? Strong answer: That the learner can map and implement a traceable RAG request flow end to end. Red flag: "It only prints an answer." Follow-up: What should the trace contain? What this proves: You value engineering artifacts, not only output.