Exact and approximate nearest-neighbor search
Exact search is the default and provides perfect recall. HNSW and IVFFlat indexes enable approximate search with a configurable speed-versus-recall tradeoff.
pgvector
Freemium
pgvector is an open-source PostgreSQL extension for storing vectors and running exact or approximate nearest-neighbor searches alongside relational data. It supports multiple vector formats, distance functions, and standard PostgreSQL operations for applications that need similarity search.
什么是 pgvector？
pgvector is an open-source PostgreSQL extension for storing vector data and performing vector similarity search. It keeps embeddings with application data in PostgreSQL rather than requiring a separate vector database, while allowing similarity queries to be combined with normal SQL operations.
The extension performs exact nearest-neighbor search by default, which provides perfect recall for the search operation. HNSW and IVFFlat indexes can be added for approximate search when lower query latency or larger collections justify trading some recall for speed. pgvector also supports quantization for workloads with many vectors.
Users interact with pgvector through PostgreSQL SQL. The extension supports vector storage, nearest-neighbor queries, distance calculations, aggregate averages, updates, deletes, upserts, bulk loading with COPY, and queries against vectors from other rows.
pgvector 能做什么？
Multiple vector representations
The extension supports single-precision, half-precision, binary, and sparse vectors, allowing storage choices to match the data and workload.
Six distance functions
SQL queries can use L2, inner product, cosine, L1, Hamming, and Jaccard distance. Hamming and Jaccard are available for binary vectors.
SQL-based data operations
Vectors can be added to new or existing tables and then inserted, updated, deleted, upserted, averaged, or loaded in bulk with PostgreSQL’s `COPY` command.
PostgreSQL-native indexing and operations
Similarity queries can be used with PostgreSQL filtering, joins, aggregates, ACID compliance, and point-in-time recovery rather than being handled outside the relational database.
Quantization and ongoing optimization
Quantization is provided for scaling collections with many vectors. The project also documents iterative index scans and performance improvements to HNSW and IVFFlat across releases.
使用场景
“Add similarity search to an existing PostgreSQL application”
Teams can add an `embedding` column to an existing table and query similar rows without moving the related records into a separate search system.
“Build semantic retrieval workflows”
Applications can store embeddings and retrieve the nearest rows using cosine, inner-product, or L2 distance, then use the returned records in an application workflow.
“Combine vector search with relational filters”
SQL queries can exclude rows, apply conditions, join related tables, or limit results while ordering by vector distance.
“Scale a larger vector collection with indexes”
When exact search becomes too slow for a workload, teams can evaluate HNSW or IVFFlat indexes and use quantization where the collection contains many vectors.
“Support varied embedding storage formats”
Projects that need reduced precision, binary representations, or sparse vectors can choose among pgvector’s supported vector types and corresponding distance operations.
常见问题
What is pgvector?
pgvector is an open-source extension that adds vector similarity search and vector data types to PostgreSQL.
How do I enable pgvector in a database?
After installing the extension, run `CREATE EXTENSION vector;` once in each database where it is needed. You can then create a column such as `embedding vector(3)` and use SQL to insert and query vectors.
Does pgvector perform exact or approximate search?
It performs exact nearest-neighbor search by default. Adding an HNSW or IVFFlat index enables approximate search, which trades some recall for speed.
Which distance functions are supported?
The documented operators support L2 distance, negative inner product, cosine distance, L1 distance, Hamming distance for binary vectors, and Jaccard distance for binary vectors.
Which PostgreSQL versions are supported?
The installation instructions state that the current release supports PostgreSQL 13 and later. The changelog also records support for PostgreSQL 18.
快速信息
- Category
- Developer Tool
- Product type
- Open-source PostgreSQL extension
- Primary function
- Vector storage and similarity search
- Query interface
- SQL through any language with a PostgreSQL client
- Vector formats
- Single-precision, half-precision, binary, and sparse
- Installation
- Source build, Docker, package managers, PostgreSQL.app, and hosted providers
pgvector 替代品
Vespa.ai
vespa.ai
Vespa is an AI search platform for building search, retrieval-augmented generation, recommendation, personalization, and agent applications over text, vectors, tensors, and structured data. It is designed for developer teams that need configurable ranking and distributed operation at production scale.
Supabase
supabase.com
Supabase is a Postgres development platform for building applications with a database, authentication, APIs, realtime features, serverless functions, storage, and vector support. It helps developers move from application code to a managed backend through a dashboard, client libraries, CLI, and generated APIs.
Upstash Vector
upstash.com
Upstash Vector is a serverless vector database for storing and querying embeddings and metadata in AI and machine-learning applications. It supports dense, sparse, and hybrid indexes through a REST API and SDKs for Python, TypeScript, Go, and PHP.
Powabase
powabase.ai
面向 AI 应用的后端平台，集成 Postgres、身份验证、存储、实时功能、检索和智能体
Weaviate
weaviate.io
Weaviate is a vector database for building AI-native applications, including retrieval-augmented generation systems. It helps developers store and retrieve vector data, connect applications to language models, and deploy the database in a managed, self-hosted, or private-cloud environment.
Spice AI
spice.ai
面向数据密集型应用和 AI 智能体的开源 SQL 查询与混合搜索引擎，零 ETL 即可运行。