Vector search

Prerequisites

Before you begin, complete the following preparations:

Install the Tablestore Go SDK and initialize a client.
A search index that contains a Vector field is created. For more information, see Create a search index.

Description

Vector search uses KnnVectorQuery to calculate distances between a query vector and vectors in a Vector field and returns the TopK most similar rows. You can use Filter to restrict candidates and MinScore to exclude rows below a score threshold.

Important

Limits apply to the number and dimensions of Vector fields and to parameters such as TopK. For more information, see Search index limits.

Note

A search index contains multiple server-side partitions. Each partition returns its own TopK nearest rows, and a coordinator merges the results. Therefore, when token-based pagination is used, the total number of retrievable rows depends on the number of index partitions.

func (client *tablestore.TableStoreClient) Search(request *tablestore.SearchRequest) (*tablestore.SearchResponse, error)

query := &search.KnnVectorQuery{
    FieldName:          "embedding",
    TopK:               proto.Int32(10),
    NumCandidates:      proto.Int32(100),
    Float32QueryVector: []float32{1, 0, 0, 0},
    MinScore:           proto.Float32(0.1),
    Filter: &search.TermQuery{
        FieldName: "category",
        Term:      "books",
    },
}
searchQuery := search.NewSearchQuery().
    SetQuery(query).
    SetSort(&search.Sort{Sorters: []search.Sorter{search.NewScoreSort()}})

response, err := client.Search(&tablestore.SearchRequest{
    TableName:   "example_table",
    IndexName:   "example_vector_index",
    SearchQuery: searchQuery,
    ColumnsToGet: &tablestore.ColumnsToGet{
        ReturnAllFromIndex: true,
    },
})
if err != nil {
    log.Fatal(err)
}

fmt.Println(response.SearchHits)

Parameters

Query request

Name	Type	Description
TableName (required)	string	The name of the data table.
IndexName (required)	string	The name of the search index.
SearchQuery (required)	search.SearchQuery	The query condition and common query configurations.
ColumnsToGet (optional)	*tablestore.ColumnsToGet	The return-column configuration. If this parameter is omitted, only primary key columns are returned.
RoutingValues (optional)	[]*tablestore.PrimaryKey	Primary key values for custom routing fields. Omit this parameter if custom routing is not configured.
TimeoutMs (optional)	*int32	The request timeout period in milliseconds.

Query configuration

Name	Type	Description
SetQuery (required)	search.Query	Specifies the query condition.
SetOffset (optional)	int32	Specifies the start position. Default value: 0. For offset-based pagination, Offset + Limit cannot exceed 100,000.
SetLimit (optional)	int32	Specifies the maximum number of rows to return. Default value: 10. Maximum value: 100. A value of 0 returns no rows.
SetCollapse (optional)	*search.Collapse	Collapses query results. For more information, see Collapse query results.
SetSort (optional)	*search.Sort	Specifies the result sort order. For more information, see Sort and paginate results.
SetGetTotalCount (optional)	bool	Specifies whether to count all matched rows. Default value: false.
SetToken (optional)	[]byte	Specifies the NextToken value returned by the previous response. This method clears Sort because the token contains the previous-page sort conditions. Do not specify Offset when you use token-based pagination.
SetSearchFilter (optional)	*search.SearchFilter	Applies a post-query filter. For more information, see Use post-query filters.
Aggregation (optional)	...search.Aggregation	Configures aggregations. For more information, see Aggregation.
GroupBy (optional)	...search.GroupBy	Configures grouping. For more information, see Aggregation.

Vector query condition

Name	Type	Description
FieldName (required)	string	The name of the Vector field.
TopK (required)	*int32	The number of nearest rows to return. A larger value generally improves recall but increases latency and cost.
Float32QueryVector (required)	[]float32	The query vector. Its dimension must match the index field configuration.
Filter (optional)	search.Query	A filter for candidate rows. All query types except vector search are supported.
MinScore (optional)	*float32	The minimum score threshold. The value must be greater than or equal to 0. Default value: 0. Only rows whose scores are greater than this value are returned.
NumCandidates (optional)	*int32	The number of candidate vectors for reranking. The value must be no less than TopK. A larger value generally improves recall but increases latency.
Weight (optional)	*float32	The relevance weight of the query condition.

Columns to return

Name	Type	Description
Columns (optional)	[]string	The attribute columns to return. This parameter takes effect only when ReturnAll and ReturnAllFromIndex are both false.
ReturnAll (optional)	bool	Specifies whether to return all attribute columns in the data table. Default value: false.
ReturnAllFromIndex (optional)	bool	Specifies whether to return all indexed attribute columns. Default value: false. Do not set this parameter and ReturnAll to true at the same time.

Response

Name	Type	Description
TotalCount	int64	The total number of matched rows. The value depends on SetGetTotalCount.
Rows	[]*tablestore.Row	The rows returned by the current query. The number does not exceed the value specified by SetLimit.
SearchHits	[]*tablestore.SearchHit	The search hits. Read this field when you use highlighting, nested inner hits, or relevance scores.
NextToken	[]byte	The token for the next page. If the value is not empty, pass it to the next query.
IsAllSuccess	bool	Indicates whether all index partitions were queried. If the value is false, partial results are returned and TotalCount may be less than the actual number of matched rows.
AggregationResults	search.AggregationResults	The aggregation results.
GroupByResults	search.GroupByResults	The grouping results.