benwtrent commented on PR #16526: URL: https://github.com/apache/lucene/pull/16526#issuecomment-5696979825
The idea makes sense. I don't have time for a deep review :(. I would prefer that this is an evolution of the HnswFiltered directly, instead of yet another way to do filtered HNSW search. Its experimental, and not the default. I think improving the algorithm is fine. One thing that troubles me, is that your benchmarks contradict ones I did. I found that Lucene's ACORN provided much better recall and latency curve vs. the plain "sweep" over a bunch of data sets, and yet your results indicate ACORN is the worst choice, for almost everything. Another interesting result I don't understand: the "positive correlation" (if it means that filter acceptance correlates positively with the nearest neighbors), I would expect almost all the filtered criteria to be much closer together as this would mean exploration over the filtered set would also more likely mean exploration of the nearest neighbors. -- 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]
