How multi-vector search changes the reranking strategy

[YYGe01]

I want to take the maximum value of multiple vector similarity matches as the final score.
Currently I only see WeightedRanker and RRFRanker.
I hope I can customize my Ranker method.

[yhmo]

I suppose this is a new feature.
What kind of ranker do you want?

[YYGe01]

I suppose this is a new feature.我想这是一个新功能。 What kind of ranker do you want? 您想要什么样的排名？

like this.
https://discuss.elastic.co/t/how-to-modify-default-scoring-calculation-when-using-multiple-knn-mixed-searches/356513
For example, I store a piece of text in three text vectors, and I want to take the maximum value of the three vectors as the final score.

[yhmo]

The link of elasic shows a weight rank of vector search and text match.
Milvus only supports rank of vector search.

"take the maximum value of the three vectors" ---------- what is "maximum value of the thress vectors"?

[YYGe01]

Official documentation includes this example.
https://milvus.io/docs/multi-vector-search.md

request_1 = AnnSearchRequest(**search_param_1) 
request_2 = AnnSearchRequest(**search_param_2) 
reqs = [request_1, request_2] 
rerank = WeightedRanker(0.8, 0.2)  

The final score should be: 0.8 * request_1_score + 0.2 * request_2_score

I would like the rerank to be:

request_1 = AnnSearchRequest(**search_param_1) 
request_2 = AnnSearchRequest(**search_param_2) 
reqs = [request_1, request_2] 
rerank = MAX(reqs)

The final score is: max(request_1_score, request_2_score)

[yhmo]

Ok, I see. I think this is a new Ranker type that can be implemented.
Wait @xiaofan-luan comment.

[yiwen92]

@YYGe01 will you ensure that request 1 and 2 are using the same metric type?

[Olimoyo]

Is this still being considered as a feature? Alternatively, can we write our own custom ranker?

[xiaofan-luan]

Is this still being considered as a feature? Alternatively, can we write our own custom ranker?

Please be patience and wait for milvus 2.6 comes out.
We have plan to support udf based reranker, and also rerank by scalar field, lke time.