hybridSearch()

This operation performs multi-vector search on a collection and returns search results after reranking.

java

public SearchResp hybridSearch(HybridSearchReq request)

Request Syntax

java

hybridSearch(HybridSearchReq.builder()
    .databaseName(String databaseName)
    .collectionName(String collectionName)
    .partitionNames(List<String> partitionNames)
    .searchRequests(List<AnnSearchReq> searchRequests)
    .topK(int topK)
    .limit(long limit)
    .outFields(List<String> outFields)
    .offset(long offset)
    .roundDecimal(int roundDecimal)
    .consistencyLevel(ConsistencyLevel consistencyLevel)
    .groupByFieldName(String groupByFieldName)
    .groupSize(Integer groupSize)
    .strictGroupSize(Boolean strictGroupSize)
    .functionScore(FunctionScore functionScore)
    .build()
);

BUILDER METHODS:

• databaseName(String databaseName)

  The name of the database. Defaults to the current database if not specified.

• clusterId(String clusterId)

  Deprecated. The ID of the cluster to query. Applies to global-cluster deployments.

• collectionName(String collectionName)

  The name of the target collection.

• partitionNames(List<String> partitionNames)

  A list of partition names to target.

• searchRequests(List<AnnSearchReq> searchRequests)

  A list of AnnSearchReq objects for hybrid search.

• topK(int topK)

  The number of top results to return.

• limit(long limit)

  The maximum number of results to return.

• outFields(List<String> outFields)

  A list of field names to include in the output.

• offset(long offset)

  The number of results to skip before returning.

• roundDecimal(int roundDecimal)

  The number of decimal places for distance/score rounding.

• consistencyLevel(ConsistencyLevel consistencyLevel)

  The consistency level for the operation.

• groupByFieldName(String groupByFieldName)

  The field name to group search results by.

• groupSize(Integer groupSize)

  The number of results to return per group.

• strictGroupSize(Boolean strictGroupSize)

  Whether to strictly enforce the group size.

• functionScore(FunctionScore functionScore)

  A FunctionScore object for custom scoring.

• ranker(CreateCollectionReq.Function ranker)

  Deprecated. A single rerank function applied to the search results. Prefer functionScore().

RETURN TYPE:

SearchResp

RETURNS:

A SearchResp object containing the merged results from all query vectors, ordered by the reranking strategy.

PARAMETERS:

• searchResults (List<List<SearchResult>>) -

  A list of search result batches, one per query vector, each containing SearchResult entries with the following fields:

  • id (Object) -

    The primary key value of the matched entity.

  • score (Float) -

    The relevance score of the match after reranking.

  • entity (Map<String,Object>) -

    A map that contains the field names and values of the matched entity.

  • primaryKey (String) -

    The name of the primary key field.

  • highlightResults (Map<String,HighlightResult>) -

    The highlight results keyed by field name, when highlighting is requested.

  • elementOffset (Long) -

    For struct-array element-level matches (via element_filter), the matched element's index within the array. Null for ordinary matches.

• sessionTs (long) -

  The session timestamp of the read.

• recalls (List<Float>) -

  The recall of each query vector, when reported.

• cost (Long) -

  The time cost of the operation.

• scannedRemoteBytes (Long) -

  The number of bytes scanned remotely during the search.

• scannedTotalBytes (Long) -

  The total number of bytes scanned during the search.

• cacheHitRatio (Float) -

  The cache hit ratio of the search.

• aggregationBuckets (List<List<AggregationBucket>>) -

  The aggregation buckets grouped by query vector, when an aggregation is requested.

EXCEPTIONS:

• MilvusClientException

  This exception will be raised when any error occurs during this operation.

Example

java

import io.milvus.v2.service.vector.request.AnnSearchReq;
import io.milvus.v2.service.vector.request.HybridSearchReq;
import io.milvus.v2.service.vector.request.FunctionScore;
import io.milvus.v2.service.vector.request.ranker.WeightedRanker;
import io.milvus.v2.service.vector.response.SearchResp;
import io.milvus.v2.common.ConsistencyLevel;

import java.util.ArrayList;
import java.util.Arrays;
import java.util.List;

// Build ANN search requests for multiple vector fields
List<AnnSearchReq> searchRequests = new ArrayList<>();
searchRequests.add(AnnSearchReq.builder()
        .vectorFieldName("float_vector")
        .vectors(floatVectors)
        .params("{\"nprobe\": 10}")
        .limit(10)
        .build());
searchRequests.add(AnnSearchReq.builder()
        .vectorFieldName("binary_vector")
        .vectors(binaryVectors)
        .limit(50)
        .build());
searchRequests.add(AnnSearchReq.builder()
        .vectorFieldName("sparse_vector")
        .vectors(sparseVectors)
        .limit(100)
        .build());

// Hybrid search with WeightedRanker via FunctionScore
SearchResp searchResp = client.hybridSearch(HybridSearchReq.builder()
        .collectionName("my_collection")
        .searchRequests(searchRequests)
        .functionScore(FunctionScore.builder()
                .addFunction(WeightedRanker.builder()
                        .weights(Arrays.asList(0.2f, 0.5f, 0.6f))
                        .build())
                .build())
        .limit(5)
        .consistencyLevel(ConsistencyLevel.BOUNDED)
        .build());

List<List<SearchResp.SearchResult>> searchResults = searchResp.getSearchResults();
for (List<SearchResp.SearchResult> results : searchResults) {
    for (SearchResp.SearchResult result : results) {
        System.out.println(result);
    }
}