Pranshu-S commented on issue #16029:
URL: https://github.com/apache/lucene/issues/16029#issuecomment-5342414708

   Right — the RaBitQ-inspired prototype I was referring to does use a 
centroid/residual representation, so it isn't data-blind. My mention of 
changing the estimator wasn't intended as an extension of the data-blind 
representation itself. I had gone down a separate path while trying to 
understand the remaining low-bit scoring loss after preconditioning, and RaBitQ 
was one of the approaches I came across there.
   
   One result that does seem directly relevant here: I tested the exact 
`HadamardRotation` from #16092 with the data-blind 1-bit BBQ path and found 
that it helps the data-blind path significantly in the cases I looked around: 
On Fashion-MNIST-784 (20K vectors, 200 held-out queries, Euclidean, HNSW 
`M=16`, `beamWidth=100`), I measured the minimum candidate count required to 
reach >=0.95 recall@10, before fp32 reranking:
   
   | Configuration | Min candidates for >=0.95 recall@10 |
   |---|---:|
   | fp32 | 10 |
   | data-blind 1-bit BBQ | 75 |
   | #16092 + data-blind 1-bit BBQ | 50 |
   
   So preconditioning reduces the search budget required at the same recall 
target from `75 -> 50`, or about **33% fewer candidates**.
   
   Removing centering still leaves a substantial gap to fp32, but the rotation 
recovers a meaningful part of that gap before introducing any residual 
representation or reranking mechanism.


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]


---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to