An odd thing is that DeepMind came out with Muvera last year (arxiv.org/html/2405.19...) which takes multi-vector and encodes back into a single vector with pretty decent results. It would have been great for that to be included here. (BM25 rules the world around us still)