[Feature]: hybrid search support choice for normalization

[yiwen92]

Is there an existing issue for this?

• I have searched the existing issues

Is your feature request related to a problem? Please describe.

Now in milvus, weighted ranker need to do normolization first, in order to combine different metric types from multi-recall ways.
However, in some embedding models like bge-m3, mgte. They do not need this normolization and they can plus the distance directly.

Describe the solution you'd like.

Add a param in rerank to control whether need do normalization or not.

Describe an alternate solution.

No response

Anything else? (Additional Context)

No response

[xiaofan-luan]

Is there an existing issue for this?

• I have searched the existing issues

Is your feature request related to a problem? Please describe.

Now in milvus, weighted ranker need to do normolization first, in order to combine different metric types from multi-recall ways. However, in some embedding models like bge-m3, mgte. They do not need this normolization and they can plus the distance directly.

Describe the solution you'd like.

Add a param in rerank to control whether need do normalization or not.

Describe an alternate solution.

No response

Anything else? (Additional Context)

No response

this means their vector distance is already normalized.

No matter what kind of normalization you do the result won't change.

There is not need to add extra complexity to the api