Question 610
Two embedding providers return vectors with different magnitudes and the team wants a consistent normalized comparison. The team is considering this step: Normalize vectors according to an evaluated similarity contract before comparison. What evidence should they review before proceeding?
Explanation
Correct Answer
Check norm type, zero-vector handling, dimensions and model assumptions. VECTOR_NORMALIZE can produce normalized values for supported SQL vector operations.
Incorrect Answers
Normalization changes magnitude, not model-space alignment; the mathematical norm and embedding compatibility both matter.
Unit normalization cannot infer direction from a zero vector or guarantee relevance; edge-case tests and the chosen similarity contract matter.
Parsing and magnitude scaling do not make incompatible dimensions or model coordinate systems comparable; compare only validated compatible vectors.