hybridSearch()

This operation conducts a hybrid search with an optional scalar filtering expression.

hybridSearch(data): Promise<SearchResults>

Request Syntax

milvusClient.hybridSearch({
    db_name: string,
    collection_name: string,
    partition_names?: string[], 
    data: HybridSearchSingleReq[], 
    limit?: number,
    offset?: number
    output_fields?: string | list[string],
    metric_type?: string,
    consistency_level?: string,
    ignore_growing?: boolean,
    rerank?: RerankerObj,
    group_by_field?: string;
    group_size?: number;
    strict_group_size?: boolean;
    hints?: string;
    round_decimal?: number;
    transformers? OutputTransformers;
    timeout?: number
})

PARAMETERS:

• db_name (string) -

  The name of the database that holds the target collection.

• collection_name (string) -

  [REQUIRED]

  The name of the collection to search.

• partition_names (string[]) -

  A list of the names of the partitions to search.

• data (HybridSearchSingleReq[]) -

  A list of search requests in the form of the HybridSearchSingleReq object.

  • data (VectorTypes[] | VectorTypes) -

    [REQUIRED]

    The query vectors. The following vector types are supported:

    • FloatVector (number[])

    • Float16Vector (number[] | Uint8Array)

    • BinaryVector (number[])

    • BFloat16Vector (number[] | Uint8Array)

    • SparseFloatVector (SparseVectorArray | SparseVectorDic | SparseVectorCSR | SparseVectorCOO)

      For details on how to generate a SparseFloatVector, refer to this code snippet.

• limit (number) -

  The total number of entities to return.

  You can use this parameter in combination with offset in param to enable pagination.

  The sum of this value and offset in param should be less than 16,384.

• offset (number) -

  The number of records to skip in the search result.

  You can use this parameter in combination with limit to enable pagination.

  The sum of this value and limit should be less than 16,384.

• output_fields (string[]) -

  A list of field names to include in each entity in return.

  The value defaults to None. If left unspecified, only the primary field is included.

• metric_type (string) -

  The metric type used to measure similarity between vectors. The value varies with the vector field type. The following table lists the mapping between vector field types and their supported metric types.

  Field Type	Dimension Range	Supported Metric Types	Default Metric Type
  FLOAT_VECTOR	2-32,768	COSINE, L2, IP	COSINE
  FLOAT16_VECTOR	2-32,768	COSINE, L2, IP	COSINE
  BFLOAT16_VECTOR	2-32,768	COSINE, L2, IP	COSINE
  INT8_VECTOR	2-32,768	COSINE, L2, IP	COSINE
  SPARSE_FLOAT_VECTOR	No need to specify the dimension.	IP, BM25 (used only for full text search)	IP
  BINARY_VECTOR	8-32,768*8	HAMMING, JACCARD, MHJACCARD	HAMMING

• consistency_level (ConsistencyLevelEnum) -

  The consistency level of the target collection. The value defaults to Bounded (1) with options of Strong (0), Bounded (1), Session (2), and Eventually (3).

• ignore_growing (boolean) -

  A boolean value indicating whether to skip the search in growing segments.

• reranker (RerankerObj) -

  A reranking strategy with its custom parameters.

  • strategy (string) -

    A re-ranking strategy. Possible values are:

    • RRF (“rrf”)

      This strategy is recommended when there is no specific emphasis. The RRF can effectively balance the importance of each vector field.

    • WEIGHTED (“weighted”)

      This strategy is recommended if you require the results to emphasize a particular vector field. The WeightedRanker allows you to assign higher weights to certain vector fields, emphasizing them more. For instance, in multimodal searches, textual descriptions of an image might be considered more important than the colors in this image.

  • params (keyValueObj) -

    The parameters are specific to reranking strategies.

    • When using the RRFRanker strategy, you need to input the parameter value k into the RRFRanker. The default value of k is 60. This parameter helps to determine how the ranks are combined from different ANN searches, aiming to balance and blend the importance across all searches.

    • When using the WeightedRanker strategy, you need to input weight values into the WeightedRanker function. The number of basic ANN searches in a Hybrid Search corresponds to the number of values that need to be inputted. The input values should be in the range [0,1], with values closer to 1 indicating greater importance.

• group_by_field (string) -

  Groups search results by a specified field to ensure diversity and avoid returning multiple results from the same group.

• group_size (number) -

  The target number of entities to return within each group in a grouping search. For example, setting group_size=2 instructs the system to return up to 2 of the most similar entities (e.g., document passages or vector representations) within each group. Without setting group_size, the system defaults to returning only 1 entity per group.

• strict_group_size (boolean) -

  This Boolean parameter dictates whether group_size should be strictly enforced. When group_size=true, the system will attempt to fill each group with exactly group_size results, as long as sufficient data exists within each group. If there is an insufficient number of entities in a group, it will return only the available entities, ensuring that groups with adequate data meet the specified group_size.

• hints (string) -

  A hints string to improve search performance.

• round_decimal (number) -

  The number of decimal places to keep in the final results.

• transformers (OutputTransformers) -

  A custom function to convert data for the following data types:

  • BFloat16Vector ((bf16bytes: Uint8Array) => BFloat16Vector;)

  • Float16Vector ((f16: Uint8Array) => Float16Vector;)

  • SparseFloatVector ((sparse: SparseVectorDic) => SparseFloatVector;)

• timeout (number) -

  The timeout duration for this operation. Setting this to None indicates that this operation timeouts when any response arrives or any error occurs.

RETURNS Promise<SearchResults>

This method returns a promise that resolves to a SearchResults object.

{
    status: object,
    results: list[string],
    recalls: list[number]
}

PARAMETERS:

• status (object) -

  • code (number) -

    A code that indicates the operation result. It remains 0 if this operation succeeds.

  • error_code (string | number) -

    An error code that indicates an occurred error. It remains Success if this operation succeeds.

  • reason (string) -

    The reason that indicates the reason for the reported error. It remains an empty string if this operation succeeds.

• results (list[object]) -

  Each result object has the following keys:

  • id (string) -

    The ID of the search result

  • score(number) -

    The similarity score of the search result.

  • Plus output fields and their values.

• recalls (list[number]) -

  Each number indicates the recall rate of a search against a query vector.

Example

const { MilvusClient, DataType } = require("@zilliz/milvus2-sdk-node")

res = await client.loadCollection({
    collection_name: "hybrid_search_collection"
})

const search = await client.search({
  collection_name: "hybrid_search_collection",
  data: [search_param_1, search_param_2],
  limit: 2,
  rerank: RRFRanker(100)
});