Good way to output embedding for search?

Right here are two options off the top of my head.

  1. Take the average of all of the output embeddings
  2. Use the CLS embedding (if it is a BERT-ish model)

This will ensure that you always have the same vector size (768, I think)