kaivalnp commented on PR #16506: URL: https://github.com/apache/lucene/pull/16506#issuecomment-5488530483
FYI there was a bug in `luceneutil` that caused the baseline exact KNN results to be incorrect for the `query-time-pre-filter` cases (see benchmark https://github.com/apache/lucene/pull/16506#issuecomment-5261243742), leading to an artificially low recall. See https://github.com/mikemccand/luceneutil/pull/614 for the bug and fix. `index-time-filter` runs are completely unaffected, and everything apart from recall for `query-time-pre-filter` is unaffected. I re-ran a [full benchmark](https://gist.github.com/kaivalnp/90a5ca5d4424e116a0505bed4cf54707) for both `main` and this PR on 100K Cohere v3 vectors, 1024 dimensions, 10K queries, across all quantization levels and vector similarities (80 combinations each). TL;DR the recall for `query-time-pre-filter` is better than reported earlier, almost on par with `index-time-filter`. Everything else (performance benefit of `index-time-filter` over `query-time-pre-filter`, and indexing / search times of the flat and de-duplicating vector formats) is unchanged. I'll merge this soon! -- 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]
