Reranker Used in Doctor-R1

[ysahil97]

Hi Authors!
Thanks for putting up the source code for Doctor-R1

I'm replicating the first two steps of the experience module of Doctor-R1.
In the paper's appendix, it is mentioned that the reranker used is BAAI/bge-reranker-base, while in the source code implementation, I see that the reranker is implemented through Cohere's api. I have a two-part question:

1. Does Cohere API support bge-reranker model calls?
2. If the answer to the previous question is negative, then, does the official implementation of Doctor R1 experience retrieval modules uses bge-reranker or Cohere API?

I would be glad if I could gain some clarity in the official implementation used.

[thu-unicorn]

To quickly answer your questions:

1. No, the official Cohere cloud API does not support bge-reranker-base.

2. For our official implementation, we hosted bge-reranker-base on our own local inference servers. Because our serving engine exposes a Cohere-compatible API endpoint, we were able to use the standard cohere Python package to query our local model simply by overriding the base url.

import cohere

# Point to the local server hosting the BGE model (see the official documentation of vllm / sglang)
base_url = "http://<your-local-server-ip>:<port>"
api_key = "your-local-token"

co = cohere.Client(base_url=base_url, api_key=api_key)

rerank = co.rerank(
    model="bge-reranker-base",
    query="What is the capital of France?",
    documents=["The capital of France is Paris", "Reranking is fun!"]
)

scores = [item.relevance_score for item in rerank.results]
print(scores)

Hope this clears up the confusion!