{
  "lesson": "67-query-rewriting-hyde",
  "title": "Query Rewriting: HyDE, Multi-Query, and Decomposition",
  "questions": [
    {
      "stage": "pre",
      "question": "What is the core mechanism of HyDE (Hypothetical Document Embeddings)?",
      "options": [
        "Embed the user query through a larger model",
        "Hash the query and look it up in a cache",
        "Run the query through multiple retrievers and pick the best one",
        "Generate a fake answer with an LLM, embed that hypothetical document, and retrieve against its vector instead of the query vector"
      ],
      "correct": 3,
      "explanation": "HyDE writes a document-shaped passage in the corpus voice and retrieves on that embedding; the query vector is replaced."
    },
    {
      "stage": "pre",
      "question": "Why is it acceptable for the LLM-generated hypothetical document in HyDE to be factually wrong?",
      "options": [
        "The reranker repairs factual errors",
        "The LLM call is rolled back when wrong",
        "The retriever cares about the token distribution, not factual correctness; the hypothetical's vector lands near the real passage",
        "The user never sees the hypothetical"
      ],
      "correct": 2,
      "explanation": "Retrieval is similarity in embedding space; if the hypothetical's vocabulary matches the corpus, the vector lands in the right region."
    },
    {
      "stage": "check",
      "question": "What is the difference between multi-query expansion and query decomposition?",
      "options": [
        "Multi-query produces N paraphrases of the same question; decomposition produces sub-questions that together cover a multi-topic question",
        "Decomposition runs locally; multi-query requires a remote model",
        "They are identical",
        "Multi-query uses cosine; decomposition uses BM25"
      ],
      "correct": 0,
      "explanation": "Paraphrases preserve intent; sub-questions split a multi-topic query into independently answerable pieces."
    },
    {
      "stage": "check",
      "question": "When does query decomposition over-split and hurt retrieval?",
      "options": [
        "When BM25 is disabled",
        "When the corpus is too small",
        "When the query is atomic (single-topic), the decomposer invents fake sub-questions and the retrievals all return the same document with reduced rank",
        "When the query is too long for the LLM"
      ],
      "correct": 2,
      "explanation": "Decomposing an atomic query splits it into fake sub-questions; their retrievals dilute the merge so the right document ranks lower."
    },
    {
      "stage": "check",
      "question": "Why do the three rewriting strategies fuse with RRF rather than score interpolation?",
      "options": [
        "Score interpolation requires per-corpus calibration; RRF combines rank-based contributions without calibration, which is required when fusing N retrievals from different rewrites",
        "Score interpolation cannot handle more than two lists",
        "RRF runs on GPUs and score interpolation does not",
        "RRF is the only operation supported by Python"
      ],
      "correct": 0,
      "explanation": "Same argument as lesson 65; rank-based fusion does not need per-corpus alpha tuning and stays stable across rewriter outputs."
    },
    {
      "stage": "post",
      "question": "What is the latency floor when running all three rewriters in parallel?",
      "options": [
        "Three LLM calls in series",
        "The slowest single retrieval",
        "One LLM call (the rewriters and retrievals run in parallel; the LLM call is the floor)",
        "Zero, because the rewriters cache everything"
      ],
      "correct": 2,
      "explanation": "All three rewriters fan out to a model call; if they run in parallel the floor is one model call latency, then the retrievals run in parallel after."
    }
  ]
}