jerry-024 commented on PR #82: URL: https://github.com/apache/paimon-vector-index/pull/82#issuecomment-5393691840
Thanks for calling out the low-`nprobe` recall trade-off. This PR intentionally follows Lance’s add-only centroid-assignment optimization: for sufficiently large centroid matrices, add uses a small graph to accelerate assignment, while query continues to select `nprobe` centroids exactly. As in Lance, this does not guarantee that an approximately assigned list is included when `nprobe` is very low. We treat this as an explicit performance/recall trade-off rather than an exact-assignment guarantee. Users that require exact coarse assignment, especially with a low `nprobe`, can disable the optimization with `approximate-assignment=false`. The default auto policy only enables it above the same `dimension × nlist` threshold used by Lance. -- 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]
