The metric is the ranking rule. Choose it before you benchmark an index. Vector search is often described as nearest-neighbor search, but nearest by what? L2 distance, cosine similarity, and inner product each reward different geometry. That difference matters when the data includes long documents, short questions, image features, product descriptions, or embeddings produced by different model families. Cosine Style Use when direction matters more than magnitude. This is common for text embeddings where the question is usually semantic closeness, not vector length. Inner Product Style Use when the model was trained or normalized for dot-product scoring, or when magnitude…
Read the full lesson
Sign up free: one personalized lesson every day, matched to your role and goals.
Already have an account? Sign in