[
https://issues.apache.org/jira/browse/LUCENE-10577?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17538474#comment-17538474
]
Robert Muir commented on LUCENE-10577:
--------------------------------------
My main concern with some custom encoding would be if it requires some slow
scalar conversion.
Currently with simple float representation you can do everything from a
float[], byte[], or mmaped data directly. See
https://issues.apache.org/jira/browse/LUCENE-9838
So if you can do stuff directly with ByteVector that would be fine. Also if you
can use "poor man's vector" with varhandles and a 64-bit long to operate on the
byte values, thats fine too. But please nothing that only works "one at a time".
> Quantize vector values
> ----------------------
>
> Key: LUCENE-10577
> URL: https://issues.apache.org/jira/browse/LUCENE-10577
> Project: Lucene - Core
> Issue Type: Improvement
> Components: core/codecs
> Reporter: Michael Sokolov
> Priority: Major
>
> The {{KnnVectorField}} api handles vectors with 4-byte floating point values.
> These fields can be used (via {{KnnVectorsReader}}) in two main ways:
> 1. The {{VectorValues}} iterator enables retrieving values
> 2. Approximate nearest -neighbor search
> The main point of this addition was to provide the search capability, and to
> support that it is not really necessary to store vectors in full precision.
> Perhaps users may also be willing to retrieve values in lower precision for
> whatever purpose those serve, if they are able to store more samples. We know
> that 8 bits is enough to provide a very near approximation to the same
> recall/performance tradeoff that is achieved with the full-precision vectors.
> I'd like to explore how we could enable 4:1 compression of these fields by
> reducing their precision.
> A few ways I can imagine this would be done:
> 1. Provide a parallel byte-oriented API. This would allow users to provide
> their data in reduced-precision format and give control over the quantization
> to them. It would have a major impact on the Lucene API surface though,
> essentially requiring us to duplicate all of the vector APIs.
> 2. Automatically quantize the stored vector data when we can. This would
> require no or perhaps very limited change to the existing API to enable the
> feature.
> I've been exploring (2), and what I find is that we can achieve very good
> recall results using dot-product similarity scoring by simple linear scaling
> + quantization of the vector values, so long as we choose the scale that
> minimizes the quantization error. Dot-product is amenable to this treatment
> since vectors are required to be unit-length when used with that similarity
> function.
> Even still there is variability in the ideal scale over different data sets.
> A good choice seems to be max(abs(min-value), abs(max-value)), but of course
> this assumes that the data set doesn't have a few outlier data points. A
> theoretical range can be obtained by 1/sqrt(dimension), but this is only
> useful when the samples are normally distributed. We could in theory
> determine the ideal scale when flushing a segment and manage this
> quantization per-segment, but then numerical error could creep in when
> merging.
> I'll post a patch/PR with an experimental setup I've been using for
> evaluation purposes. It is pretty self-contained and simple, but has some
> drawbacks that need to be addressed:
> 1. No automated mechanism for determining quantization scale (it's a constant
> that I have been playing with)
> 2. Converts from byte/float when computing dot-product instead of directly
> computing on byte values
> I'd like to get people's feedback on the approach and whether in general we
> should think about doing this compression under the hood, or expose a
> byte-oriented API. Whatever we do I think a 4:1 compression ratio is pretty
> compelling and we should pursue something.
--
This message was sent by Atlassian Jira
(v8.20.7#820007)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
