Score Reranker udf?

[JackTan25]

Hello, I'm trying to use milvus's multi-column top-k result. But I find the recall is very low. Finally, I find the reason lies that milvus use https://milvus.io/docs/reranking.md#Weighted-Scoring-WeightedRanker, which will not use distance as score. I hope we can use the origin distance as the score computation, that's (w0 * L2(vector0, v_const0) + w1 * L2(vector1,v_const1) + w2 * L2(vector2,v_const2)) by using WeightRanker(w0,w1,w2). Hope add new ranker to implement this one.

[yhmo]

Wait @czs007 to comment.

[czs007]

@JackTan25 please refer to this code.

milvus/internal/proxy/reScorer.go

Line 84 in 648d566

func (ws *weightedScorer) getActivateFunc() activateFunc {

Milvus currently does not support passing UDF (User-Defined Function) during ranking.
The quickest way to make modifications is to add a parameter to the WeightRanker and switch back to the previous calculation method.
If you are in a hurry, you can directly modify the code here.

[JackTan25]

thanks @czs007 I will modify the source code and test it locally

[JackTan25]

@czs007 I modify only one row code, but after test, I find out the recall is zero. That's very strange. Maybe my modification is wrong??

[JackTan25]

bug here: #33294