[Bug]: Potential Index Mismatch in Rerank Model Results for bge-reranker-v2-m3

[1006076811]

Is there an existing issue for the same bug?

• I have checked the existing issues.

Branch name

main

Commit ID

last

Other environment information

No response

Actual behavior

t seems there may be a logical issue in the rerank model when calling the bge-reranker-v2-m3 model. Specifically, the result includes an index field that marks the pre-sorted index of the items. However, in the similarity function, the code directly returns the following array:

np.array([d["relevance_score"] for d in res["results"]])

This array appears to have lost the index information, which can lead to a mismatch between the relevance scores and the original indexes. I'm unsure if I misunderstood something, but it seems like a bug, as the scores and the indexes do not align correctly.

Expected behavior

The relevance scores returned should maintain a proper mapping to their corresponding pre-sorted indexes.

Steps to reproduce

Call the bge-reranker-v2-m3 model using the rerank feature.

Observe that the index field is included in the results but not preserved when returning the relevance scores.

Check for the mismatch between the scores and their corresponding indexes.

Additional information