[
https://issues.apache.org/jira/browse/LUCENE-10069?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17404687#comment-17404687
]
Michael Sokolov commented on LUCENE-10069:
------------------------------------------
I think that because we prune neighbors to maintain a graph with a maximum
fanout for each node, we can actually end up with a disconnected graph _ I saw
this when testing with very small graphs; I originally had an
assertGraphConnected method, but had to stop using it for tests after adding
pruning. Even if the graph is not disconnected (larger graphs probably aren't),
I think there can be "backwaters" where almost we never go?
> HNSW can miss results with very large k
> ---------------------------------------
>
> Key: LUCENE-10069
> URL: https://issues.apache.org/jira/browse/LUCENE-10069
> Project: Lucene - Core
> Issue Type: Bug
> Reporter: Julie Tibshirani
> Priority: Minor
>
> Performing a kNN search with very large k (where k is close to the total
> number of live docs) can return fewer than k documents. This occurs because
> searches aren't guaranteed to be able to visit every node in the graph.
> Specifically, when choosing entry points for the search, we make k random
> draws _with replacement_, and sometimes end up with fewer than k entry
> points. These entry points may not provide a connection to all other nodes in
> the graph.
> This is an unusual case, but I think it'd be nice if we could always return k
> docs when they are available (or at least document the behavior?) We're
> currently working on adding graph layers (LUCENE-10054), which will change
> how entry points are selected. Maybe we can revisit this issue once that work
> is further along to see if a fix is still needed.
> Here's an example test failure showing the problem. We happen to select
> {{numDocs=101}} and {{k=100}}.
> {code:java}
> ./gradlew test --tests TestKnnVectorQuery.testRandom
> -Dtests.seed=3B6CE0E105431E18
> {code}
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org
For additional commands, e-mail: issues-h...@lucene.apache.org
