[Bug]: [hybrid_search] The rerank effect needs more improvements when setting different metric type for different vector field using "WeightedRanker" reranker

[binbinlv]

Is there an existing issue for this?

• I have searched the existing issues

Environment

- Milvus version: master-latest
- Deployment mode(standalone or cluster): both
- MQ type(rocksmq, pulsar or kafka):    all
- SDK version(e.g. pymilvus v2.0.0rc2): 2.4.0rc57
- OS(Ubuntu or CentOS): 
- CPU/Memory: 
- GPU: 
- Others:

Current Behavior

The rerank effect may be bad when setting different metric type for different vector field using "WeightedRanker" reranker:

for example:

when setting metric type "COSINE" for float vector field A, and setting "L2" for float vector field B, then now it will choose the metric type for the first vector field in schema as the sorted way for the hybrid search result.

But this way seems not reflecting the similarity after reranking, because "COSINE" is the larger the similar, and "L2" is the smaller the similar.

And another point is that it seems not very meaningful to weighted sum of two values in huge range difference, just like the range of "COSINE" is "-1 ~ 1", and "L2" is "-∞ ～ +∞“.

Expected Behavior

A better WeightedRanker algorithm design which could reflect the real similarity

Steps To Reproduce

1. create a collection with 3 vector fields
2. insert data
3. create index in "COSINE", "L2", "IP" metric type in the 3 vector fields
4. load
5. hybrid search using "WeightedRanker" reranker

    @pytest.mark.parametrize("primary_field", [ct.default_int64_field_name])
    def test_hybrid_search_different_metric_type_each_field(self, primary_field, dim, auto_id, is_flush,
                                                 enable_dynamic_field, metric_type):
        """
        target: test hybrid search for fields with different metric type
        method: create connection, collection, insert and search
        expected: hybrid search successfully with limit(topK)
        """
        # 1. initialize collection with data
        collection_w, _, _, insert_ids, time_stamp = \
            self.init_collection_general(prefix, True, auto_id=auto_id, dim=dim, is_flush=is_flush, is_index=False,
                                         primary_field=primary_field,
                                         enable_dynamic_field=False, multiple_dim_array=[dim, dim])[0:5]
        # 2. extract vector field name
        vector_name_list = cf.extract_vector_field_name_list(collection_w)
        vector_name_list.append(ct.default_float_vec_field_name)
        log.debug(vector_name_list)
        flat_index = {"index_type": "FLAT", "params": {}, "metric_type": "L2"}
        collection_w.create_index(vector_name_list[0], flat_index)
        flat_index = {"index_type": "FLAT", "params": {}, "metric_type": "IP"}
        collection_w.create_index(vector_name_list[1], flat_index)
        flat_index = {"index_type": "FLAT", "params": {}, "metric_type": "COSINE"}
        collection_w.create_index(vector_name_list[2], flat_index)
        collection_w.load()
        # 3. prepare search params
        req_list = []
        search_param = {
            "data": [[random.random() for _ in range(dim)] for _ in range(1)],
            "anns_field": vector_name_list[0],
            "param": {"metric_type": "L2", "offset": 0},
            "limit": default_limit,
            "expr": "int64 > 0"}
        req = AnnSearchRequest(**search_param)
        req_list.append(req)
        search_param = {
            "data": [[random.random() for _ in range(dim)] for _ in range(1)],
            "anns_field": vector_name_list[1],
            "param": {"metric_type": "IP", "offset": 0},
            "limit": default_limit,
            "expr": "int64 > 0"}
        req = AnnSearchRequest(**search_param)
        req_list.append(req)
        search_param = {
            "data": [[random.random() for _ in range(dim)] for _ in range(1)],
            "anns_field": vector_name_list[2],
            "param": {"metric_type": "COSINE", "offset": 0},
            "limit": default_limit,
            "expr": "int64 > 0"}
        req = AnnSearchRequest(**search_param)
        req_list.append(req)
        # 4. hybrid search
        hybrid_search = collection_w.hybrid_search(req_list, WeightedRanker(0.1, 0.9, 1), default_limit,
                                                   check_task=CheckTasks.check_search_results,
                                                   check_items={"nq": 1,
                                                                "ids": insert_ids,
                                                                "limit": default_limit})[0]
        log.debug(hybrid_search[0].ids)
        log.debug(hybrid_search[0].distances)

Milvus Log

No response

Anything else?

No response

[yanliang567]

/unassign

[czs007]

working on it

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.
Rotten issues close after 30d of inactivity. Reopen the issue with /reopen.