[
https://issues.apache.org/jira/browse/SOLR-18581?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Adam Quigley updated SOLR-18581:
--------------------------------
Description:
[https://issues.apache.org/jira/browse/SOLR-18267] added knnAlgorithm="flat"
which disables HNSW graph construction for brute-force vector search use cases.
DenseVectorFields currently reject the \{!knn}, \{!knn_text_to_vector} and
\{!vectorSimilarity} query parsers and only support vectorSimilarity() function
queries for flat fields. ScalarQuantizedDenseVectorFields support all KNN query
parsers for FLOAT32 vectors for flat fields.
This limitation exists because Lucene99FlatVectorsFormat used by
DenseVectorField doesn't implement search(), so the KNN query parsers would
silently return no results.
Lucene104ScalarQuantizedVectorsFormat used by ScalarQuantizedDenseVectorField
with knnAlgorithm="flat" implements search by brute-force scanning vectors.
A solution likely requires either a Lucene change or a brute-force search in
the reader of Solr101FlatVectorFormat.
was:
[https://issues.apache.org/jira/browse/SOLR-18267] added knnAlgorithm="flat"
which disables HNSW graph construction for brute-force vector search use cases.
DenseVectorFields currently reject the \{!knn}, \{!knn_text_to_vector} and
\{!vectorSimilarity} query parsers and only support vectorSimilarity() function
queries for flat fields. ScalarQuantizedDenseVectorFields support all KNN query
parsers for FLOAT32 vectors for flat fields.
This limitation exists because Lucene99FlatVectorsFormat used by
DenseVectorField doesn't implement search(), so the KNN query parsers would
silently return no results.
Lucene104ScalarQuantizedVectorsFormat used by ScalarQuantizedDenseVectorField
with knnAlgorithm="flat" implements search by brute-force scanning FLOAT32
vectors.
A solution likely requires either a Lucene change or a brute-force search in
the reader of Solr101FlatVectorFormat.
> Support KNN query parsers for DenseVectorField with knnAlgorithm="flat"
> -----------------------------------------------------------------------
>
> Key: SOLR-18581
> URL: https://issues.apache.org/jira/browse/SOLR-18581
> Project: Solr
> Issue Type: New Feature
> Components: vector-search
> Affects Versions: 10.1, 9.12
> Reporter: Adam Quigley
> Priority: Minor
>
> [https://issues.apache.org/jira/browse/SOLR-18267] added knnAlgorithm="flat"
> which disables HNSW graph construction for brute-force vector search use
> cases.
> DenseVectorFields currently reject the \{!knn}, \{!knn_text_to_vector} and
> \{!vectorSimilarity} query parsers and only support vectorSimilarity()
> function queries for flat fields. ScalarQuantizedDenseVectorFields support
> all KNN query parsers for FLOAT32 vectors for flat fields.
> This limitation exists because Lucene99FlatVectorsFormat used by
> DenseVectorField doesn't implement search(), so the KNN query parsers would
> silently return no results.
> Lucene104ScalarQuantizedVectorsFormat used by ScalarQuantizedDenseVectorField
> with knnAlgorithm="flat" implements search by brute-force scanning vectors.
>
> A solution likely requires either a Lucene change or a brute-force search in
> the reader of Solr101FlatVectorFormat.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
