{
  "lesson": "12-video-understanding-pipeline",
  "title": "Capstone 12 — Video Understanding Pipeline (Scene, QA, Search)",
  "questions": [
    {
      "stage": "pre",
      "question": "Why is a scene-level index still needed even when a long-context VLM can read a 2-hour video natively?",
      "options": [
        "Scene cuts are the only thing a VLM understands",
        "Vector indexes do not support video files",
        "Long-context VLMs cannot stream output",
        "Ingesting 100 hours of video into a queryable corpus needs scene-level retrieval, even when individual videos can be read whole"
      ],
      "correct": 3,
      "explanation": ""
    },
    {
      "stage": "pre",
      "question": "Which three vector types does each scene get in the multi-vector index?",
      "options": [
        "Audio waveform, mel-spectrogram, and MFCCs",
        "Pose, depth, and optical flow",
        "Hash, gzip, and CRC",
        "Caption embedding, keyframe embedding, and transcript embedding"
      ],
      "correct": 3,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "How are results from the three retrieval streams merged at query time?",
      "options": [
        "Via reciprocal rank fusion across the three ranked lists",
        "By averaging cosine similarities",
        "By picking the caption hit only",
        "By taking the union without scoring"
      ],
      "correct": 0,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "What does the temporal grounding step refine inside the top scene?",
      "options": [
        "The (start, end) timestamp window that contains the answer",
        "The keyframe embedding dimensionality",
        "The transcript word timestamps from Whisper",
        "The number of vectors stored in Qdrant"
      ],
      "correct": 0,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "Which question class is reported separately because it is a known hallucination hotspot?",
      "options": [
        "Translation questions",
        "Speaker identification questions",
        "Counting and action-type questions, where VLMs miscount or mis-order events",
        "Descriptive questions about scenery"
      ],
      "correct": 2,
      "explanation": ""
    },
    {
      "stage": "post",
      "question": "Which grounding metric does the rubric measure on a held-out set?",
      "options": [
        "Perplexity",
        "Temporal-grounding intersection-over-union (IoU)",
        "Acceptance rate",
        "MRR@10"
      ],
      "correct": 1,
      "explanation": ""
    },
    {
      "stage": "post",
      "question": "Why does the pipeline require cited timestamps in answers?",
      "options": [
        "Timestamps are required by Whisper output",
        "Timestamps are required by Qdrant payloads",
        "They reduce the embedding dimension",
        "They let the viewer jump to the exact (video_id, start, end) so users can verify the claim"
      ],
      "correct": 3,
      "explanation": ""
    }
  ]
}