Puneet Ahuja created SOLR-17990:
-----------------------------------
Summary: FAISS integration for Solr vector search
Key: SOLR-17990
URL: https://issues.apache.org/jira/browse/SOLR-17990
Project: Solr
Issue Type: Task
Reporter: Puneet Ahuja
This issue proposes to add integration of
[FAISS|https://github.com/facebookresearch/faiss] (Facebook AI Similarity
Search) into solr, as it offers different indexing strategies that can be
better for some workloads. Lucene 11.0 (unreleased main branch) includes FAISS
support via FaissKnnVectorsFormat, but Solr uses Lucene 10.3.1, which doesn't
have it. Since FaissKnnVectorsFormat is only in Lucene's unreleased main branch
(targeted for Lucene 11.0), we need to backport the FAISS-related code to make
it available for current Solr.
This requires backporting the FAISS codec from Lucene 11.0 to work with Lucene
10.3.1, creating a codec factory to integrate it, and using Java's Foreign
Function & Memory API to call the native FAISS library. Users should be able to
choose FAISS per field by setting knnAlgorithm="faiss" in their schema.xml,
while keeping existing HNSW setups working. Because this backports unreleased
Lucene code, it should go in the solr-sandbox repository rather than the main
Solr repository, following the pattern for experimental or pre-release
integrations.
This would give Solr users access to FAISS capabilities and more flexibility
for vector search, while staying compatible with current functionality.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
