← back to post

With impact of vector magnitudes on similarity score, are there scenarios where using dot-product might accidentally exaggerate similarity, like for vectors with large magnitudes?
Or are web content/passage embeddings too few dimensions to worry about that?

Brian Crouch · Questions · · Jun 19, 14:45
1 reply

Absolutely. If direction is more important than intensity, use cosine similarity or normalize embeddings before computing dot-product.

Dan Petrovic · SuggestsExpands · · Jun 20, 13:17

Sign in with Google to reply.