Binary maxsim via hamming

[michaelbridge]

Binary quantization and hamming distance are critical for scaling multi-vector representations (i.e., Colbert).

It looks as though hamming for binary vectors has already been implemented.

While a hamming-based maxsim can be implemented over this with a postgres function per approach here, is this something that might be supported/optimized within the library?

Beyond this, is an unpack_bits operation to convert a binary vector into a float representation (to improve accuracy in a subsequent rerank step) something contemplated?

[michaelbridge]

Ah, looks like RaBitQ is an implementation of binary quant, so perhaps this question is better restated as, how can Colbert/Colpali late interaction be optimized within this framework?

[gaocegege]

Hey, maybe you could checkout https://blog.vectorchord.ai/supercharge-vector-search-with-colbert-rerank-in-postgresql

[michaelbridge]

Thanks, but I linked that blog post above. Without a multi-vector index, that doesn't scale.

[VoVAllen]

We're baking some approach at #197 based on https://github.com/jlscheerer/xtr-warp/tree/main/warp/search. Please stay tuned!