mccullocht commented on PR #16200: URL: https://github.com/apache/lucene/pull/16200#issuecomment-4624943905
> True, a normalized byte vector implies normalization before quantization. The "normalization" might be as simple as the linear scaling of the magnitude is the same for all input vectors. I would guess that in many cases the input float vector was normalized before quantization. Aside from the linear scaling being uniform making other assumptions seems brittle. There's also #12263 which notes a problem with dot similarity and `byte[]` vectors compressing the score range. My guess is the input vectors in this case were normalized before the model quantized them. I think it would be very difficult to change the behavior of byte/dot similarity even if it makes sense -- it would likely be disruptive for downstream consumers of Lucene. -- 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]
