[Bug] LlamaIndex embeddings using wrong method

[logan-markewich]

The llamaindex tea embeddings are using a private method

GenAIComps/comps/embeddings/llama_index/embedding_tei.py

Line 21 in cd83854

embed_vector = embeddings._get_query_embedding(input.text)

It's not clear to me if the embedding service is meant to handle query embeddings or document embeddings. But either way, we should be using embed_model.get_text_embedding(text) or embed_model.get_query_embedding(query)

Likely we should have two endpoints, one for query and one for normal text documents.

We might also want to consider using get_text_embeddings_batch() instead of processing one document at a time, but again, depends on how we want to define our embedding endpoints

[srinarayan-srikanthan]

the object is used to call the functions, but agree that we might want to consider text_embedding_batch()

[logan-markewich]

Right, but I think we also don't want to ignore the fact that its pretty common for embedding models to embed queries and documents differently (cohere and nomic are both examples that need this)

Maybe this needs to be a parameter on the request?

[logan-markewich]

I've also noticed that the embeddings API in general only works by sending a single piece of text at a time. This is neither efficient nor openai compatible sadly

Related PR: run-llama/llama_index#16666