[
https://issues.apache.org/jira/browse/SOLR-18424?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18112987#comment-18112987
]
Kevin Liang commented on SOLR-18424:
------------------------------------
Another consideration regarding performance is how the .vec random access gets
memory mapped. We may want to avoid filling up the page cache (quantization is
used to mitigate this problem in the first place), as it may cause thrashing +
bad assumptions of the OS to prefetch more of the file. An optimization Elastic
has done is to have a vector reader that deliberately does not memory map what
it pulls. I don't know if there is a better pattern, but for the purposes of
this ticket might just be out of scope for first pass of this feature.
> Add native rerank for dense vectors
> -----------------------------------
>
> Key: SOLR-18424
> URL: https://issues.apache.org/jira/browse/SOLR-18424
> Project: Solr
> Issue Type: Improvement
> Reporter: Kevin Liang
> Priority: Major
>
> Solr 10 has introduced support for binary quantized dense vectors
> (SOLR-17781), however binary quantized vectors by themselves have a pretty
> noticeable recall hit (only achieving ~60% recall compared to non-quantized
> KNN by some benchmarks). The solution is to pair binary quantization with
> oversampling + rerank phase. While it is possible to do this currently with
> Solr's generic rerank function through something like:
> {code:java}
> q={!knn f=vector_quantized topK=100}[...]&rq={!rerank reRankDocs=100
> reRankQuery=$rrq reRankWeight=1.0
> reRankOperator=replace}&rrq={!func}vectorSimilarity(FLOAT32, DOT_PRODUCT,
> vector_quantized, [...])&rows=10{code}
> this is verbose, cumbersome, and easy to get subtley wrong behavior. Propose
> that we support reranking directly in the knn query parser via an
> "oversample" factor param such as
> {code:java}
> q={!knn f=vector_quantized topK=10 oversample=10}[0.12, 0.34, ...]{code}
> "oversample" here meaning retrieve n * k documents, rerank with raw vectors,
> and return top k.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
