int4 doesn't handle cosine similarity correctly

[benwtrent]

Description

just with some default settings, glove-200 does poorly with int4 HNSW and when using cosine. The bug occurs on merge. When recalculating the quantiles, the vectors aren't normalized like they should, so the quantiles will get all out of whack. We can actually see this in some of the below experiments. All these values should indicate a normalized vectory, however some of the results are > 1 or < -1.

Some experiments I have done:

Dynamic confidence interval:

• Floats used to build the graph (e.g. when buffer is large enough): 0.693 (I regrettably didn't get the quantiles here...)
• When int4 vectors are used 0.189 (minQuantile=-1.2686034, maxQuantile=1.3250866)

0.5 confidence interval (locally patched to allow it):

• int4 vectors building the graph 0.649 (minQuantile=-0.277436, maxQuantile=0.29298353)

0.75 confidence interval (locally patched to allow it):

• int4 vectors building the graph 0.535 (minQuantile=-0.48500386, maxQuantile=0.5028781)

0.9 confidence interval:

• int4 vectors building the graph 0.407 (minQuantile=-0.71112806, maxQuantile=0.73441404)

[benwtrent]

Actually, this might be a bug. Looking at the code, I am not sure we normalize the vectors when building the quantizer & using cosine.

[benwtrent]

Yep, not normalizing during quantile merging is the actual bug. I have a local patch that fixes this. Need to write up some tests & will push up a change soon.

Good find @naveentatikonda!