feat: support sparse text embeddings

[joshua-mo-143]

• I have looked for existing issues (including closed) about this

Feature Request

Following on from #268, we should try to support sparse text model embeddings.

This will enable users to carry out things like hybrid search for RAG.

Motivation

Sparse embeddings are quite good for some types of NLP and information search.

They are also quite efficient and small which makes them good storage-wise.

Proposal

Unfortunately, Fastembed does not support sparse text models by default despite apparently having it in the library - you have to create your own.

This issue however, points to the bm42-rs library which does support it. So we can implement it this way if we are OK with adding another library. (we'll likely need to do some exploratory work to figure this part out)

Alternatives

Not sure?