HybridSearch()

Executes multiple vector searches and combines them with the requested reranking strategy.

pub async fn hybrid_search(&self, request: HybridSearchRequest) -> Result<SearchResponse>

Request Syntax

let request = HybridSearchRequest::builder()
    .collection_name("books")
    .sub_requests(vec![
        SubSearchRequest::builder()
            .vector_field("embedding")
            .vectors(SearchVectors::Float(vec![vec![0.1, 0.2, 0.3, 0.4]]))
            .metric_type(MetricType::Cosine)
            .build()?,
    ])
    .rerank(RRFRerank::new())
    .limit(5)
    .build()?;

REQUEST FIELDS:

• database_name: Option<String>

  Name of the target database; uses the default database when empty.

• collection_name: String

  Name of the collection to search. Required.

• partition_names: Vec<String>

  Partition names to restrict the search to.

• sub_requests: Vec<SubSearchRequest>

  Child vector searches, each with its own vector field, query vectors, filter, and metric. At least one sub-search is required.

• rerank: Option<Function>

  Reranking strategy applied to the combined results, such as RRF or weighted reranking. Use RRFRerank::new() or WeightedRerank::new().

• limit: i64

  Number of final hits returned after reranking. Defaults to 10.

• offset: i64

  Number of leading hits to skip. Defaults to 0.

• round_decimal: i64

  Number of decimal places for the returned scores; -1 disables rounding. Defaults to -1.

• ignore_growing: bool

  Whether to ignore growing segments during the search. Defaults to false.

• extra_params: HashMap<String, String>

  Additional reranking parameters.

• group_by_field: String

  Scalar field used to group the combined results.

• group_size: i64

  Number of results to return per group. Defaults to 1.

• strict_group_size: bool

  Whether each group must return exactly group_size results. Defaults to false.

• output_fields: Vec<String>

  Names of the scalar fields returned for each hit.

• consistency_level: Option<ConsistencyLevel>

  Consistency guarantee for the search; the collection default is used when unset. See ConsistencyLevel.

SubSearchRequest

Each entry in sub_requests describes one vector search against a single vector field. Build it with SubSearchRequest::builder() and pass the list to HybridSearchRequest::builder().sub_requests(...).

REQUEST FIELDS:

• vector_field: String

  Name of the vector field to search. Required.

• vectors: SearchVectors

  Query vectors, wrapped in the SearchVectors enum matching the field’s data type (e.g. SearchVectors::Float, SearchVectors::Sparse). Required. See SearchVectors.

• filter: String

  Scalar filter expression applied to the entities before search.

• filter_templates: HashMap<String, FilterTemplateValue>

  Placeholder values referenced by the filter expression. See FilterTemplateValue.

• limit: i64

  Maximum number of hits returned for this sub-search. Defaults to 0.

• metric_type: Option<MetricType>

  Metric type for measuring vector similarity; falls back to the index metric when unset. See MetricType.

• extra_params: HashMap<String, String>

  Index-specific search parameters, such as nprobe for IVF indexes.

• radius: Option<f64>

  Similarity radius used with range_filter for range search; matches results whose distance is within the range.

• range_filter: Option<f64>

  Lower/upper similarity bound that works with radius for range search.

• timezone: String

  Timezone applied when filtering on Timestamptz fields.

RETURN TYPE:

Result

RETURNS:

A SearchResponse containing the ranked hits for each query vector. Returns an Error if the operation fails.

SearchResponse
    results() // -> &SearchResults
    session_timestamp() // -> u64
    cost() // -> i64
    scanned_remote_bytes() // -> i64
    scanned_total_bytes() // -> i64
    cache_hit_ratio() // -> f32

METHODS:

• results() (&SearchResults) -

  The ranked hits, one group per query vector.

  • get_results() (&[SingleResult]) -

    The hits, one per query vector.

    • get_ids() (&Ids) -

      The IDs of the hits for one query vector. See Ids.

    • get_scores() (&[f32]) -

      The similarity scores of the hits.

    • get_element_indices() (Option<&[i64]>) -

      The element indices of the hits, when present.

    • get_output_fields() (&[FieldData]) -

      The output field values of the hits. See FieldData.

    • get_output_field_names() (&[String]) -

      The names of the output fields.

    • get_primary_field_name() (&str) -

      The name of the primary-key field.

    • get_score_field_name() (&str) -

      The name of the score field.

    • get_highlight_results() (&[HashMap]) -

      The text-highlighting results per hit, when full-text search highlighting was requested.

      • get_field_name() (&str) -

        The name of the highlighted field.

      • get_fragments() (&[String]) -

        The highlighted text fragments.

      • get_scores() (&[f32]) -

        The relevance scores of the fragments.

    • len() (usize) -

      The number of hits for one query vector.

    • is_empty() (bool) -

      Whether no hits were returned for the query vector.

    • get_output_field() (Option<&FieldData>) -

      The value of the named output field for the hit, when present. See FieldData.

    • get_output_rows() (Result>) -

      The hits as entity rows. See EntityRow.

    • get_output_row() (Result) -

      The hit at the given index as an entity row. See EntityRow.

  • get_recalls() (&[f32]) -

    The recall rate per query vector.

  • len() (usize) -

    The number of hit groups returned.

  • is_empty() (bool) -

    Whether no hits were returned.

  • get_agg_buckets() (&[Vec]) -

    The aggregation buckets produced by the search.

    • get_key() (&[BucketKeyEntry]) -

      The bucket keys, one entry per grouped field.

      • get_field_id() (i64) -

        The ID of the field the key belongs to.

      • get_field_name() (&str) -

        The name of the field the key belongs to.

      • get_value() (&AggregationBucketValue) -

        The value of the bucket key.

    • get_count() (i64) -

      The number of rows in the bucket.

    • get_metrics() (&HashMap) -

      The metric values computed for the bucket. See AggregationMetricValue.

    • get_hits() (&[AggregationHit]) -

      The representative hits of the bucket.

      • get_pk() (Option<&AggregationHitPk>) -

        The primary key of the representative hit.

      • get_score() (f32) -

        The score of the representative hit.

      • get_fields() (&[AggregationHitField]) -

        The field values of the representative hit.

        • get_field_id() (i64) -

          The ID of the field.

        • get_field_name() (&str) -

          The name of the field.

        • get_value() (&AggregationHitFieldValue) -

          The value of the field.

    • get_sub_groups() (&[AggregationBucket]) -

      The nested sub-buckets of the bucket.

• session_timestamp() (u64) -

  The session timestamp of the search.

• cost() (i64) -

  The execution cost reported by the server.

• scanned_remote_bytes() (i64) -

  The number of bytes scanned from remote storage.

• scanned_total_bytes() (i64) -

  The total number of bytes scanned.

• cache_hit_ratio() (f32) -

  The ratio of scanned bytes served from the cache.

Example

let request = HybridSearchRequest::builder()
    .collection_name("books")
    .sub_requests(vec![
        SubSearchRequest::builder()
            .vector_field("embedding")
            .vectors(SearchVectors::Float(vec![vec![0.1, 0.2, 0.3, 0.4]]))
            .metric_type(MetricType::Cosine)
            .build()?,
    ])
    .rerank(RRFRerank::new())
    .limit(5)
    .build()?;
let resp = client.hybrid_search(request).await?;