Create Vector Index

Syntax

CREATE TABLE table_name( columns_difinition, INDEX index_name (column_name) USING VECTOR PROPERTIES( "property1" = "value1", "property2" = "value2" ) );

columns_difinition: Define the field information of the table, the last field must be separated by a comma

INDEX: Keyword

index_name: Custom name for the index

column_name: The name of the field that needs to be indexed

VECTOR: Keyword, indicating vector index

COMMENT: Specify the description information of the index

PROPERTIES: Specify the parameters of the vector INDEX

Parameter Name	Optional Values	Default Value	Remarks
distance.function	l2_distancecosine_distancenegative_dot_productjaccard_distancehamming_distance	cosine_distance	For convenience, please use negative_dot_product for dot product scenarios
scalar.type	f32, f16, i8, b1	f32	The type of vector elements in the vector index, which can be different from the vector column
m	Recommended not to exceed 1000	16	Maximum number of neighbors in the HNSW algorithm
ef.construction	Recommended not to exceed 5000	128	Size of the candidate set when constructing the index in the HNSW algorithm
reuse.vector.column	false/true	false	Whether to reuse the data of the vector column to save storage space
compress.codec	uncompressed/zstd/lz4	uncompressed	Compression algorithm for the vector index; not effective when reusing the column
compress.level	fatest/default/best	default	Compression algorithm level
compress.byte.stream.split	false/true	true	Rearrange float bits before compression
compress.block.size	Integer greater than 1M	16777216	Compression block size
conversion.rule	default/as_bits	default	When indexing vector(tinyint, N) type by bits, use as_bits

Vector Index Scalar Type and Column Type

Index element type refers to the type specified by scalar.type in properties; when the types are inconsistent, certain conversion rules (mostly cast) will be applied, except when using type b1.

Index Element Type	Supported Vector Element Types	Remarks
b1	tinyint, int, float	When "conversion.rule" = "as_bits", each bit in vector(tinyint, N) is treated as an element in the vector. When "conversion.rule" = "default", the vector is binarized.
i8	tinyint, int, float	When the vector element type is inconsistent with the index, cast is performed (note that overflow may occur).
f16	int, float	When the vector element type is inconsistent with the index, cast is performed.
f32	int, float	When the vector element type is inconsistent with the index, cast is performed.

Examples

CREATE TABLE test_vector1 ( vec vector(float, 4), id int， index test_vector1_vec_idx (vec) using vector properties ( "scalar.type" = "f32", "distance.function" = "l2_distance" ) )

Reference Documentation

• Build Index
• Drop Index
• List All Indexes
• View Index Details

User Guide

Vector Index User Guide

Add Vector Index to Existing Table

Syntax

CREATE VECTOR INDEX [IF NOT EXISTS] index_name ON TABLE [schema].table_name(col_name) PROPERTIES( "property1" = "value1", "property2" = "value2" )

VECTOR: Index type, vector index

index_name: Index name, located under the schema of the target table. Index names cannot be duplicated within the same schema. The index must be in the same schema as the target table.

PROPERTIES: Specify parameters for INDEX

Description

Executing CREATE INDEX is only effective for new data. Existing data requires BUILD INDEX to build the vector index.

The index must be in the same schema as the target table; creating an index across schemas will result in an error.

Example

CREATE TABLE test_vector2 ( vec vector(float, 512), id int ) ; CREATE vector INDEX test_vector2_vec_idx on (vec) properties ( "scalar.type" = "f32", "distance.function" = "l2_distance" );