pgvector is an open source vector similarity search tool for PostgreSQL. It allows you to store vector data alongside other data types in a database and provides the following features and functionalities:
Single-precision, half-precision, binary, and sparse vectors.
Mainstream distance measures including Manhattan distance (L1), Euclidean distance (L2), inner product, cosine distance, Hamming distance, and Jaccard distance.
Exact and approximate nearest neighbor search.
Mainstream vector approximate similarity search indexes such as Hierarchical Navigable Small World (HNSW) and Inverted File with Flat Compression (IVFFlat).
Various PostgreSQL client languages.
This extension also supports powerful PostgreSQL features, such as atomicity, consistency, isolation, and durability (ACID) compliance, point-in-time recovery, and JOINs.
Copyright notice
PolarDB for PostgreSQL incorporates the pgvector extension. The pgvector extension, including its documentation and functionality, was originally developed and maintained by the PostgreSQL Global Development Group. Alibaba Cloud has made modifications to adapt pgvector for its proprietary vector database solution.
Copyright (c) 1996-2025, PostgreSQL Global Development Group
Single-precision vectors
Each single-precision vector occupies 4 * dimensions + 8 bytes of storage space. Each element of the vector is a single-precision floating-point number (similar to the real type in PostgreSQL), and all elements must be valid and finite values (cannot be NaN, Infinity, or -Infinity). Single-precision vectors can have up to 16,000 dimensions.
Operators
Operator | Description | Version requirement |
| Element-wise addition | None |
| Element-wise subtraction | None |
| Element-wise multiplication | 0.5.0 or later |
| Concatenation | 0.7.0 or later |
| Euclidean distance | None |
| Negative inner product | None |
| Cosine distance | None |
| Manhattan distance | 0.7.0 or later |
Functions
Function | Description | Version requirement |
| Quantizes a vector into binary format. | 0.7.0 or later |
| Calculates the cosine distance between vectors. | None |
| Calculates the inner product of vectors. | None |
| Calculates the Manhattan distance between vectors. | 0.5.0 or later |
| Calculates the Euclidean distance between vectors. | None |
| Normalizes a vector by using its Euclidean norm | 0.7.0 or later |
| Extracts a subvector from a given vector. | 0.7.0 or later |
| Returns the number of dimensions of a vector. | None |
| Calculates the Euclidean norm of a vector. | None |
Aggregate function | Description | Version requirement |
| Calculates the average of vectors. | None |
| Calculates the sum of vectors. | 0.5.0 or later |
Half-precision vectors
Each half-precision vector occupies 2 * dimensions + 8 bytes of storage space. Each element of the vector is a half-precision floating-point number, and all elements must be valid and finite values (cannot be NaN, Infinity, or -Infinity). Half-precision vectors can have up to 16,000 dimensions.
Operators
Operator | Description | Version requirement |
| Element-wise addition | 0.7.0 or later |
| Element-wise subtraction | 0.7.0 or later |
| Element-wise multiplication | 0.7.0 or later |
| Concatenation | 0.7.0 or later |
| Euclidean distance | 0.7.0 or later |
| Negative inner product | 0.7.0 or later |
| Cosine distance | 0.7.0 or later |
| Manhattan distance | 0.7.0 or later |
Functions
Function | Description | Version requirement |
| Quantizes a half-precision vector into binary format. | 0.7.0 or later |
| Calculates the cosine distance between half-precision vectors. | 0.7.0 or later |
| Calculates the inner product of half-precision vectors. | 0.7.0 or later |
| Calculates the Manhattan distance between half-precision vectors. | 0.7.0 or later |
| Calculates the Euclidean distance between half-precision vectors. | 0.7.0 or later |
| Calculates the Euclidean norm of a half-precision vector. | 0.7.0 or later |
| Normalizes a half-precision vector by using its Euclidean norm. | 0.7.0 or later |
| Extracts a subvector from a given half-precision vector. | 0.7.0 or later |
| Returns the number of dimensions of a half-precision vector. | 0.7.0 or later |
Aggregate function | Description | Version requirement |
| Calculates the average of vectors. | 0.7.0 or later |
| Calculates the sum of vectors. | 0.7.0 or later |
Binary vectors
Each binary vector occupies dimensions / 8 + 8 bytes of storage space. For more information, see Bit String Types.
Operators
Operator | Description | Version requirement |
| Hamming distance | 0.7.0 or later |
| Jaccard distance | 0.7.0 or later |
Functions
Function | Description | Version requirement |
| Calculates the Hamming distance between binary vectors. | 0.7.0 or later |
| Calculates the Jaccard distance between binary vectors. | 0.7.0 or later |
Sparse vectors
Each sparse vector occupies 8 * number of non-zero elements + 16 bytes of storage space. Each element of the vector is a single-precision floating-point number, and all elements must be valid and finite values (cannot be NaN, Infinity, or -Infinity). Sparse vectors can have up to 16,000 non-zero elements.
Operators
Operator | Description | Version requirement |
| Euclidean distance | 0.7.0 or later |
| Negative inner product | 0.7.0 or later |
| Cosine distance | 0.7.0 or later |
| Manhattan distance | 0.7.0 or later |
Functions
Function | Description | Version requirement |
| Calculates the cosine distance between sparse vectors. | 0.7.0 or later |
| Calculates the inner product of sparse vectors. | 0.7.0 or later |
| Calculates the Manhattan distance between sparse vectors. | 0.7.0 or later |
| Calculates the Euclidean distance between sparse vectors. | 0.7.0 or later |
| Calculates the Euclidean norm of a sparse vector. | 0.7.0 or later |
| Normalizes a sparse vector by using its Euclidean norm. | 0.7.0 or later |