{
  "lesson": "64-chunking-strategies-advanced",
  "title": "Chunking Strategies, Compared",
  "questions": [
    {
      "stage": "pre",
      "question": "Why is the chunking choice an upper bound on retrieval quality and not just a hyperparameter?",
      "options": [
        "Because the LLM never reads chunks",
        "Because BM25 has no role in modern RAG",
        "Because no embedding, reranker, or LLM downstream can recover an answer span that was cut into incoherent pieces",
        "Because the embedding model is fixed at training time"
      ],
      "correct": 2,
      "explanation": "If the chunk that holds the answer span is incoherent the embedding moves to the wrong cluster and downstream components cannot repair the damage."
    },
    {
      "stage": "pre",
      "question": "Which strategy is the strongest single-strategy default when you have no prior knowledge of the corpus?",
      "options": [
        "Sentence packer",
        "Fixed-window",
        "Recursive split",
        "Structural markdown"
      ],
      "correct": 2,
      "explanation": "Recursive split adapts per region by trying the strongest separator first and falling back, so it survives inconsistent structure."
    },
    {
      "stage": "check",
      "question": "What does recall@k measure in this lesson?",
      "options": [
        "Whether the top-k chunks contain the literal query string",
        "The cosine similarity between the query and the top-k chunks",
        "The number of unique documents in the top-k",
        "Whether any of the top-k chunks overlaps the gold answer span"
      ],
      "correct": 3,
      "explanation": "Each query carries gold offsets; recall@k is 1 if any top-k chunk overlaps the gold span, else 0; averaged across queries."
    },
    {
      "stage": "check",
      "question": "Why does fixed-window chunking fail badly on code or YAML corpora?",
      "options": [
        "It cuts identifiers and key-value pairs in half, producing chunks that embed to noise",
        "It loses Unicode characters",
        "It cannot be parallelized",
        "It produces chunks larger than the embedding model's context window"
      ],
      "correct": 0,
      "explanation": "Character-level cuts inside a symbol destroy the meaning the embedding model is supposed to capture."
    },
    {
      "stage": "check",
      "question": "What is the centroid drift termination criterion in semantic chunking?",
      "options": [
        "Cosine similarity between the running mean vector and the next sentence drops below a threshold",
        "Character count exceeds the max",
        "Sentence count exceeds a hardcoded limit",
        "The chunker has consumed the entire embedding budget"
      ],
      "correct": 0,
      "explanation": "Semantic chunking cuts when the running centroid no longer matches the next sentence by a configured similarity threshold."
    },
    {
      "stage": "post",
      "question": "What stale-embedding failure mode does semantic chunking introduce that the other four strategies do not?",
      "options": [
        "Chunks too small for the embedding context",
        "Chunk boundaries change whenever you change the chunk-model, requiring a coordinated rebuild",
        "Chunks are no longer indexable by BM25",
        "Chunks lose their start and end offsets"
      ],
      "correct": 1,
      "explanation": "Semantic clustering depends on an embedding model; swapping the model changes the boundaries, so the chunk-model must be pinned separately or rebuilt together with the retrieval model."
    }
  ]
}