Reranker API “top_k” Support

[etwk]

Feature request

Reran top_n is widely supported, for example here: https://github.com/run-llama/llama_index/blob/main/llama-index-integrations/postprocessor/llama-index-postprocessor-jinaai-rerank/llama_index/postprocessor/jinaai_rerank/base.py

Suggest adding this feature. Thanks.

Motivation

This is a key feature in reranking at this moment.

Your contribution

I could submit a PR, please let me know what files need to be updated.

[michaelfeil]

I think you are looking for rerankiner models? And yes, these are supported under the Rerank endpoint. View the Readme.md.

A top_k argument is not widley supported, and there are no openAI rerank specs.

[michaelfeil]

https://docs.cohere.com/reference/rerank is supported - why would you need it? I personally see that you can do this in two lines of code client side.

[etwk]

https://docs.cohere.com/reference/rerank is supported - why would you need it? I personally see that you can do this in two lines of code client side.

Maybe when a 3rd party app does not support API that does not support selecting the top n, and we have no direct control over how to use the output from reranker.

[michaelfeil]

@etwk Your instructions are not clear.

Please deploy, e.g. https://huggingface.co/mixedbread-ai/mxbai-rerank-xsmall-v1 via

infinity_emb v2 —model-id mixedbread-ai/mxbai-rerank-xsmall-v1

[michaelfeil]

@aloababa

Can you align this implementation to one of the following API protocols?

https://docs.voyageai.com/reference/reranker-api {return_documents: bool = False} {.., documents: str}
https://jina.ai/reranker/ {return_documents: bool = False} -> {.., documents: str}
https://docs.cohere.com/reference/rerank {no_kwarg_for_documents} -> {document: {text: str}}
https://huggingface.github.io/text-embeddings-inference {return_text: bool = False} -> {.., text: str}