Closed cydrain closed 1 year ago
/kind improvement
This PR removes IVF Range Search nprobe
parameter and searches until no results are found.
My concern is that the radius might result in many buckets only having a few or even one point within the radius. Thus the performance will have a vast degradation. Is there any performance report?
Checked the performance report internally. I'm not sure whether the performance gain is from pulling up the bitset check or the algorithm itself.
The UX here changes, the user can no longer adjust their QPS/Recall at the fixed radius. Is this the way we want to go down? @liliu-z
Checked the performance report internally. I'm not sure whether the performance gain is from pulling up the bitset check or the algorithm itself.
The UX here changes, the user can no longer adjust their QPS/Recall at the fixed radius. Is this the way we want to go down? @liliu-z
If necessary, what we should do it to expose another set of APIs with definition that allows users to balance recall/QPS instead of mixing them.
This PR removes IVF Range Search
nprobe
parameter and searches until no results are found.My concern is that the radius might result in many buckets only having a few or even one point within the radius. Thus the performance will have a vast degradation. Is there any performance report?
https://zilliverse.feishu.cn/docx/JMVMd91KbodZKqxs2TrcIWywn5F BTW, I think we need to align the definition of interface first. Then we can see how to balance recall/QPS by modifying the tolerance.
[APPROVALNOTIFIER] This PR is APPROVED
This pull-request has been approved by: cydrain, liliu-z
The full list of commands accepted by this bot can be found here.
The pull request process is described here
@cydrain Please associate the related issue to the body of your Pull Request. (eg. “issue: #”)