Skip to content

Investigate BBQ quantizing centroids & rescoring #131234

@benwtrent

Description

@benwtrent

Description

Digging into #129950 its tricky to get absolutely correct.

As a sibling issue to improve centroid vector scoring (which could be done in addition to indexing the centroids), we should consider quantizing centroids to single bit, and reranking with 4 bit quantization.

When the number of centroids gets large, I would expect this to have a pretty significant impact.

Some things to consider:

  • Smallest centroid quantization size should be dictated by the smallest in postings list. Meaning if postings list are 2 bit, centroids should be 2 bit
  • We will definitely need to oversample and rescore. We can do something simple and do 3x oversampling by default for single bit and something cleverer later.
  • The initial bit quantizations should be 1 bit and 4 bit. Though, once we get to 2bit, I think we should just jump to 7bit for higher fidelity.

Metadata

Metadata

Assignees

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions