mccullocht commented on issue #16029: URL: https://github.com/apache/lucene/issues/16029#issuecomment-5344921381
We had a similar discussion about rotation and centering on a PR for [TurboQuant](https://github.com/apache/lucene/pull/15903#issuecomment-4200627888). Quantizers often perform better with a Gaussian component distribution -- RaBitQ and TurboQuant lean on this particularly heavily. If you don't _need_ to be data blind, centering is still tremendously helpful. RaBitQ also has a notion of statistically bounded error on distance estimates, [which I think can be adapted for OSQ](https://github.com/apache/lucene/issues/15789). There's a potential adaptive loop here: I use the bounds to figure out how many vectors I should rerank using a higher fidelity representation of the vector (up to f32) and that might help feed selection of quantization parameters including bit rate, rotation, or even truncation for Matryoshka Representation Learning. -- 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]
