{
  "questions": [
    {
      "stage": "pre",
      "question": "For a visual search product where queries are text and gallery items are images, which backbone do you pick first?",
      "options": ["A CLIP or SigLIP model — they learn a shared text-image embedding space so cosine similarity between text query and image gallery is meaningful out of the box", "A supervised ResNet-50", "A handcrafted ORB descriptor", "A self-supervised DINOv2 model"],
      "correct": 0,
      "explanation": "Retrieval with text queries requires embeddings where text and image end up in the same space. CLIP/SigLIP train exactly this with the contrastive loss in Lesson 18. Supervised ImageNet features are image-only. DINOv2 is strong but image-only. Text queries demand a vision-language encoder."
    },
    {
      "stage": "pre",
      "question": "What does 'semi-hard mining' mean in triplet-loss training?",
      "options": ["For each anchor, select a negative that is further than the positive but still within the margin; easy negatives contribute no gradient and hardest negatives can destabilise training, so semi-hard is the sweet spot", "Mining cryptocurrency for training examples", "Using only the hardest possible negatives", "Picking random triplets"],
      "correct": 0,
      "explanation": "Triplet loss gradient is non-zero only when d(a, n) < d(a, p) + margin. Easy negatives (already far) give zero gradient. Hardest negatives can collapse training. Semi-hard — d(a, p) < d(a, n) < d(a, p) + margin — is where the loss is informative without being unstable. FaceNet introduced this recipe in 2015 and it is still the default for triplet fine-tuning."
    },
    {
      "stage": "post",
      "question": "Your retrieval system reports recall@10 = 0.95 but recall@1 = 0.42. What should you conclude?",
      "options": ["The model is overfitting", "Data is leaking from gallery to queries", "The evaluation is broken", "The embedding space has the correct structure — relevant items are usually in the top 10 — but ranking within the top 10 is noisy; a re-ranking model (cross-encoder, second-stage scorer) or longer metric-learning fine-tune will fix recall@1 without damaging recall@10"],
      "correct": 3,
      "explanation": "recall@K vs @1 tells you where the information sits. A high recall@10 with low recall@1 means retrieval is approximately right but the model is not confident about ordering the top few. Re-ranking — a separate model that scores every top-K candidate against the query — is the production fix and is what search engines like Pinterest and Google Photos use."
    },
    {
      "stage": "post",
      "question": "Cosine similarity on unnormalised embeddings is not cosine similarity — what goes wrong?",
      "options": ["The matmul overflows", "It mixes direction with magnitude; cos(a, b) is only correct when both vectors are L2-normalised. On unnormalised embeddings, large-norm items dominate regardless of direction, skewing ranks toward vectors with big magnitudes", "PyTorch raises an error", "Nothing, it is fine"],
      "correct": 1,
      "explanation": "cos(a, b) = (a . b) / (||a|| * ||b||). Skipping the normalisation gives you raw inner product, which ranks by direction weighted by magnitude. On typical neural embeddings magnitudes vary 2-3x across samples, and that alone dominates the ranking. Always L2-normalise before computing cosine or before indexing in FAISS IndexFlatIP."
    },
    {
      "stage": "post",
      "question": "Between instance-level retrieval (find this exact car) and category-level retrieval (find cars), which requires metric-learning fine-tuning?",
      "options": ["Instance-level. Off-the-shelf DINOv2 and CLIP embeddings discriminate between categories well but not between visually similar instances of the same category. A triplet or contrastive fine-tune on (same instance, different instance) pairs tightens the embedding around each instance", "Both require the same amount of fine-tuning", "Category-level", "Neither"],
      "correct": 0,
      "explanation": "Category-level retrieval is what pretrained models already do: cats cluster near cats, dogs near dogs. Instance-level — same car, same face, same product SKU — needs metric learning because two instances of the same product look very similar in the general embedding space. Fine-tuning with triplet or InfoNCE on instance-labelled pairs shrinks intra-instance distance and grows inter-instance distance."
    }
  ]
}