Feature Request: Can the embeddings endpoint with llama.cpp server generate sparse vectors using models like bge-me that support dense/sparse embeddings

[nirmalavhcl]

Prerequisites

• I am running the latest code. Mention the version if possible as well.
• I carefully followed the README.md.
• I searched using keywords relevant to my issue to make sure that I am creating a new issue that is not already open (or closed).
• I reviewed the Discussions, and have a new and useful enhancement to share.

Feature Description

https://huggingface.co/BAAI/bge-m3 is an embedding model that supports both dense and sparse embeddings/retrieval. I can use llama-server /v1/embeddings endpoint to generate dense embeddings. Can llama-server optionally generate sparse vectors using BM25 lexical score?

Motivation

Enhancement to use a single model to support all bge-m3 capabilities

Possible Implementation

No response

[ggerganov]

I'm not familiar with the concept of "sparse embeddings". Can you explain how it works?

[nirmalavhcl]

sparse vector/embeddings is akin to fulltext search. Its used along with regular/dense embeddings for Hybrid searches in Vector db.

These embeddings are mostly BM25(TF-IDF based).

https://qdrant.tech/articles/sparse-vectors/