About the specific implementation of muvera

General
1

I read the paper MUVERA: Multi-Vector Retrieval via Fixed Dimensional Encodings, and I reproduced it myself in C++ and Python. I was also surprised to find that weaviate also implemented this paper.

But I found that weaviate’s recall rate is much higher than my code implementation and other code implementations on github. I encode the multi-vector into a single vector and add it to the vector index such as HNSW. I want to know whether weaviate has made other optimizations besides the solution in the paper.

2

hi @ee-wizard !!

Welcome to our community :hugs:

Have you seen our blog post about it? Here:

Codewise, this is probably a good starting point:

Probably my colleague Roberto Esposito will be able to go in deep about it.

Also, consider joining our public slack if you want to discuss this further.

Let me know if this helps!

Thanks!