KeyBERTInspired representation with precomputed embeddings

Nobody has claimed this yet.

I am having some trouble to do some create some fine tuned representations. I am using OpenAI ada embeddings, I precalculate them and use it as backend. However, when I tried to use KeyBERTInspired for finetuning it try to create again some embeddings.

If I don't use the OpenAI backend the problem is that it cannot use the embeddings already generated and try to create new.

However, if I put the openAI backend, the problem is for empty tokens that send them to the API and as it cannot create teh embedding an error is returned.

Is there a way to use prior calculated embeddings of documents or how can we fill the empty documents in this case? Thank you so much in advance.

Start by reproducing the reported behavior with KeyBERTInspired, precomputed document embeddings, and the OpenAI backend. Trace how representations handle precomputed embeddings and empty tokens; the issue is done when both workflows avoid unnecessary embedding requests and no longer fail on empty documents.

Written by the indexing model from the issue text.