{
  "lesson": "22-embedding-models-deep-dive",
  "title": "Embedding Models — The 2026 Deep Dive",
  "questions": [
    {
      "stage": "pre",
      "question": "What is a dense embedding?",
      "options": [
        "One fixed-size vector per text where cosine similarity ranks passages by semantic proximity",
        "A token-level alignment matrix",
        "A sparse weight per vocab token",
        "A graph over documents"
      ],
      "correct": 0,
      "explanation": "Dense embeddings give one vector per text (typically 384-3072 dim); cosine ranks similarity."
    },
    {
      "stage": "pre",
      "question": "What does Matryoshka Representation Learning enable?",
      "options": [
        "Faster softmax",
        "Multilingual training",
        "Cross-encoder rescoring",
        "Truncating a trained embedding to its first N dimensions and getting a still-useful smaller embedding"
      ],
      "correct": 3,
      "explanation": "Matryoshka training makes the first N dims of the vector standalone-useful, enabling cheap truncation."
    },
    {
      "stage": "check",
      "question": "How do multi-vector (ColBERT) embeddings score query-doc pairs?",
      "options": [
        "Cosine of mean-pooled token vectors",
        "Cross-entropy",
        "MaxSim: for each query token find the most similar document token, then sum the maxima",
        "Earth-mover's distance"
      ],
      "correct": 2,
      "explanation": "ColBERT-style late interaction uses MaxSim across per-token vectors."
    },
    {
      "stage": "check",
      "question": "What does BGE-M3 output simultaneously?",
      "options": [
        "Only a sparse vector",
        "Only a dense vector",
        "Dense, sparse, and multi-vector (colbert) representations from one model in a single inference",
        "Only a colbert vector"
      ],
      "correct": 2,
      "explanation": "BGE-M3 emits three retrieval modes from one model, useful for fused hybrid scoring."
    },
    {
      "stage": "check",
      "question": "Why must you re-normalize a Matryoshka-truncated vector before cosine similarity?",
      "options": [
        "Truncation changes the vector norm; without re-normalizing, dot product no longer equals cosine",
        "Cosine ignores normalization",
        "Truncation breaks training",
        "Required by FAISS only"
      ],
      "correct": 0,
      "explanation": "Re-normalizing after truncation restores unit norm so dot product equals cosine again."
    },
    {
      "stage": "post",
      "question": "Why do BGE models often need a query-side prefix string?",
      "options": [
        "To compress queries",
        "Prefixes change tokenizer behavior",
        "Required by FAISS",
        "BGE was trained with an explicit query prompt; omitting it costs 3-5 points recall"
      ],
      "correct": 3,
      "explanation": "BGE models expect a 'Represent this sentence for searching...' prefix on queries."
    },
    {
      "stage": "post",
      "question": "Why is MTEB necessary but not sufficient for picking an embedding model?",
      "options": [
        "MTEB only covers English",
        "MTEB rewards bigger models",
        "MTEB ignores latency",
        "Leaderboard ranks are average across many tasks; your specific domain may differ, so always benchmark on your data"
      ],
      "correct": 3,
      "explanation": "MTEB averages across tasks; domain-specific eval can flip the ranking."
    },
    {
      "stage": "post",
      "question": "When should you add SPLADE sparse retrieval alongside dense embeddings?",
      "options": [
        "On small corpora only",
        "Whenever the encoder is multilingual",
        "When queries are keyword-heavy or contain identifiers/codes that dense embeddings blur",
        "Whenever Matryoshka is used"
      ],
      "correct": 2,
      "explanation": "SPLADE captures lexical/keyword matches that dense models can miss; fuse with dense via RRF."
    }
  ]
}