Reduce dimension of colbert vector

#47
by feliciaAI - opened

I'm trying out the colbert vectors at the moment and I'm quite happy with the performance. I would however, if possible, need to reduce the dimension size of the embeddings and reshape them to 2-dim matrices(= one array/row for each embedded chunk to be clear). I'm aware that this will most likely affect the performance negatively, but I would still like to try if a solution like this is better than the one I've used before.

Anyone who's tried something similar?

Another approach to the question is if anyone knows a nice way for storage and retrieval of multi dimensional vectors?

@feliciaAI you can use vector databases like Datastax or Mongodb also released support for storing and indexing vectors.

Sign up or log in to comment