benwtrent commented on PR #13566: URL: https://github.com/apache/lucene/pull/13566#issuecomment-2275704679
> So this patch tries to ensure that every node on each level is reachable from one of the entry points of that level (which are just the nodes on the next higher level). This is easy to compute and seems to correspond to the property we would like to guarantee: no orphaned nodes. This is a very nice exploitation of HNSW. One concern I have is local minima. Since we only ever use a single entry point, might be that we hit a region that is actually disconnected from the true nearest neighbors. So, no matter what we do (no matter how much we increase efSearch), we will never get the true nearest neighbors. What do we think of increasing the number of entry points to the lower level? > The results do show that this approach does a good (better than before) job of linking about disconnected components. You have to look hard to see a measurable impact on recall. Latency and indexing time don't seem to be affected I would imagine recall is helped given abusive edge cases. -- 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
