[DOC] Faiss Scalar Quantization FP16 (SQfp16)

[naveentatikonda]

What do you want to do?

• Request a change to existing documentation
• Add new documentation
• Report a technical problem with the documentation
• Other

Tell us about your request. Provide a summary of the request and all versions that are affected.
In k-NN plugin we mainly support vectors of type float where each dimension is 32 bits. This is getting expensive for use cases that requires ingestion on a large scale where we need to construct, load, save and search graphs(for native engines nmslib and faiss) which is getting even more costlier. Even though we have the byte vector support, it only supports lucene engine and also there is a considerable reduction in the recall when compared to float 32.

Adding support for Faiss SQFP16 helps to reduce the memory and storage footprints without compromising on recall where when user provides the 32 bit float vectors, the Faiss engine quantizes the vector into FP16 using their scalar quantizer (users don’t need to do any quantization on their end), stores it and decodes it back to FP32 while returning the results during search operations.

This feature is expected to be launched in OS version 2.12. So, it will be supported in 2.12 and future versions.

What other resources are available? Provide links to related issues, POCs, steps for testing, etc.
opensearch-project/k-NN#1138
opensearch-project/k-NN#830

[naveentatikonda]

Please add V2.11 label

[naveentatikonda]

Please change the label to v2.12

[hdhalter]

@naveentatikonda - Is this still on track for 2.12? Thanks!

[naveentatikonda]

@naveentatikonda - Is this still on track for 2.12? Thanks!

Yes @hdhalter this feature should be ready for 2.12. We are just waiting on the Faiss maintainers to merge our changes. Thanks!
facebookresearch/faiss#3166