msokolov commented on PR #14823: URL: https://github.com/apache/lucene/pull/14823#issuecomment-3006267677
The right default value for beam width is a little hard to say. Values reported in public vary wildly. Try asking Gemini what is the right default and it refuses to be pinned down! Unfortunately it depends - on the application setup (does it have a fancy second-stage ranker, or do we rely on the vector scores directly for ranking), on the vector data set (its dimensions, the size of the index). We're finding that "high quality" vectors might require lower values for these exploration parameters, and applications that gather lots of semantic hits and then later re-rank them might be less sensitive to lower recall and can tolerate cheaper indexing settings. You can see that these sources recommend values like 16, 64, 128, and 256: * https://opensearch.org/blog/a-practical-guide-to-selecting-hnsw-hyperparameters/ * https://www.pinecone.io/learn/series/faiss/hnsw/ Maybe a good rule of thumb would be to pick the place where the slope of these latency/recall curves is 45 degrees? But of course that will depend on the units of the chart. My view is this is something that, at least as a default, we could detune a bit. Maybe we could retreat to 64? But anyway this issue is about merge policy segment counts I think so we should probably open a new issue if we want to change that -- 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: issues-unsubscr...@lucene.apache.org For queries about this service, please contact Infrastructure at: us...@infra.apache.org --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org For additional commands, e-mail: issues-h...@lucene.apache.org
