Practice Bank: Embedding Retrieval Scenarios
Use starter.py to classify when dense, sparse, or hybrid retrieval should be used.
Scenario 1: Password reset vs credential rotation
Semantic match with different words.
Scenario 2: Error code AZ-4032-B
Exact token match matters more than meaning.
Scenario 3: Stale policy vector
High similarity but inactive metadata.
Scenario 4: Product SKUs
Dense retrieval may blur identifiers.
Scenario 5: Support paraphrase
User language differs from article title.
Scenario 6: Legal clause number
Hybrid retrieval protects exact references.
Scenario 7: FAQ duplicate
Vectors help cluster similar answers.
Scenario 8: Multi-language question
Embedding model capability matters.
Scenario 9: Sensitive HR content
Provider choice and privacy review matter.
Scenario 10: Re-index after model change
Embedding version controls release safety.