[Feature]: Enable `/score` endpoint for all embedding models

[maxdebayser]

🚀 The feature, motivation and pitch

Currently only cross-encoder models support the /score endpoint. But it would make sense to enable it also for the embedding models using bi-encoding, i.e. calculating a cosine similarity score between the embedding vectors.

cc: @DarkLight1337

Alternatives

No response

Additional context

No response

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[DarkLight1337]

cc @flaviabeo

[DarkLight1337]

Not sure why I can't assign you, so I've pinged you instead.

[flaviabeo]

Ok! I will work on this!

[noooop]

refer to ColBERT

Calculating the cosine similarity score between embedding vectors on the server side is very inefficient

I think it is more reasonable to only open the /score endpoint to models with "BCD" architecture.

[DarkLight1337]

I think it is more reasonable to only open the /score endpoint to models with "BCD" architecture.

We can easily support batched inference in Score API so that (a) can also be done efficiently. That is, given queries Q and documents D, we compute their embeddings separately, and then return the score for each pair between Q and D. I think sentence-transformers has similar API.

[noooop]

We can easily support batched inference in Score API so that (a) can also be done efficiently. That is, given queries Q and documents D, we compute their embeddings separately, and then return the score for each pair between Q and D. I think sentence-transformers has similar API.

Okay, although this is not a typical application scenario of the A architecture

[DarkLight1337]

In production, the document embeddings should be pre-computed and indexed for better efficiency. Nevertheless, this API can provide a convenient way to quickly experiment with different inputs.

[noooop]

I still think /score should not support architecture A to avoid api abuse.