向量新贵 VectorChord 吊打 pgvector、milvus、es 等
P
digoal约 5 分钟
digoal约 5 分钟向量新贵 VectorChord 吊打 pgvector、milvus、es 等
向量数据库的痛点
高维稠密向量, 1百万条往上, 会遇到几重瓶颈 建立索引慢 查询速率变慢 召回率变低 消耗内存急剧增加
VectorChord 使用如下技术很好的解决了以上问题, 达到的效果超过市面上的专业向量数据库
量化, 减少计算量, 32bit -> 1bit ranker阶段时按32bit精度计算 ivf 分区 支持GPU, 支持外部build索引, 兼容PostgreSQL数据文件, 可直接导入. 建议先生成ivf的中心点集群, 需要少量时间, 例如几十万条之后再build index. 效果更佳
VectorChord比对milvus, pgvector, es等产品的数据详见
https://blog.vectorchord.ai/vectorchord-store-400k-vectors-for-1-in-postgresql
除了支持传统的向量检索, VectorChord还支持“向量数组类型”的检索
vec arr @# vec query arr
该功能的场景假设:
一段大文本按语义分片后成为多个小段/句子, 每个小段/句子一个向量? 查询时, 将一个query分成若干个观点进行查询, 每个观点一个向量?
下面使用PolarDB 15 开源版本, 测试一下VectorChord
PolarDB 15 + VectorChord
测试环境使用:
《维基百科(wikipedia) RAG 优化 | PolarDB + AI》
部署rust
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
echo". \"\$HOME/.cargo/env\"" >> ~/.bashrc
. ~/.bashrc
$ rustc --version
rustc 1.86.0 (05f9846f8 2025-03-31)
配置国内镜像
echo"
[source.crates-io]
replace-with = 'ustc'
[source.ustc]
registry = \"sparse+https://mirrors.ustc.edu.cn/crates.io-index/\"
" >> /home/postgres/.cargo/config.toml
下载VectorChord插件代码
cd /data
git clone --depth 1 -b 0.3.0 https://github.com/tensorchord/VectorChord
cd /data/VectorChord
编译安装VectorChord插件依赖环境
cd /data/VectorChord
# pgrx 版本请参考不同版本的 Cargo.toml 文件
cargo install --locked --version 0.13.1 cargo-pgrx
cargo pgrx init # create PGRX_HOME 后, 立即ctrl^c 退出
cargo pgrx init --pg15=`which pg_config` # 不用管报警
Validating /home/postgres/tmp_polardb_pg_15_base/bin/pg_config
Initializing data directory at /home/postgres/.pgrx/data-15
安装clang-16 + (依赖)
cd /data
wget https://apt.llvm.org/llvm.sh
sudo apt update
sudo apt install lsb-release wget software-properties-common gnupg
sudo chmod +x llvm.sh
sudo ./llvm.sh
编译安装VectorChord插件
cd /data/VectorChord
export CC=/usr/bin/clang-19 # Replace with the actual path
export CXX=/usr/bin/clang++-19 # Replace with the actual path
PGRX_IGNORE_RUST_VERSIONS=y cargo pgrx install --release --pg-config `which pg_config`
配置postgresql.conf
vi ~/primary/postgresql.conf
shared_preload_libraries='vchord,pg_jieba,pg_bigm,pgml,$libdir/polar_vfs,$libdir/polar_worker'
重启PolarDB for postgresql
pg_ctl restart -m fast -D ~/primary
在数据库中安装插件
postgres=# create extension vchord;
CREATE EXTENSION
postgres=# \dx
List of installed extensions
Name | Version | Schema | Description
---------------------+---------+------------+---------------------------------------------------------------------------------------------
http | 1.6 | public | HTTP client for PostgreSQL, allows web page retrieval inside the database.
openai | 1.0 | public | OpenAI client.
pg_bigm | 1.2 | public | text similarity measurement and index searching based on bigrams
pg_bulkload | 3.1.22 | public | pg_bulkload is a high speed data loading utility for PostgreSQL
pg_jieba | 1.1.0 | public | a parser for full-text search of Chinese
pg_stat_statements | 1.10 | public | track planning and execution statistics of all SQL statements executed
pg_trgm | 1.6 | public | text similarity measurement and index searching based on trigrams
pgml | 2.10.0 | pgml | Machine Learning and AI functions from postgresml.org
pgstattuple | 1.5 | public | show tuple-level statistics
plpgsql | 1.0 | pg_catalog | PL/pgSQL procedural language
plpython3u | 1.0 | pg_catalog | PL/Python3U untrusted procedural language
polar_feature_utils | 1.0 | pg_catalog | PolarDB feature utilization
polar_vfs | 1.0 | public | polar virtual file system for different storage
sslinfo | 1.2 | public | information about SSL certificates
vchord | 0.3.0 | public | vchord: Vector database plugin for Postgres, written in Rust, specifically designed for LLM
vector | 0.8.0 | public | vector data type and ivfflat and hnsw access methods
(16 rows)
使用举例:
CREATE TABLE items (id bigserial PRIMARY KEY, embedding vector(3));
INSERT INTO items (embedding) SELECT ARRAY[random(), random(), random()]::real[] FROM generate_series(1, 1000);
CREATE INDEX ON items USING vchordrq (embedding vector_l2_ops) WITH (options = $$
residual_quantization = true
[build.internal]
lists = []
$$);
SET vchordrq.probes TO '';
SELECT * FROM items ORDER BY embedding <-> '[3,1,2]' LIMIT 5;
id | embedding
-----+-----------------------------------
228 | [0.9514653,0.82165444,0.98645663]
762 | [0.98221064,0.8921851,0.882103]
157 | [0.97980326,0.5610302,0.96323115]
963 | [0.98450637,0.5665642,0.9378539]
98 | [0.93809456,0.83065367,0.9353343]
(5 rows)
vec数组
drop table items;
CREATE TABLE items (embedding vector(3));
INSERT INTO items (embedding) SELECT ARRAY[random(), random(), random()]::real[] FROM generate_series(1, 1000);
CREATE INDEX ON items USING vchordrq (embedding vector_l2_ops) WITH (options = $$
residual_quantization = true
[build.internal]
lists = [1000]
spherical_centroids = false
$$);
SET vchordrq.probes = 10;
SELECT * FROM items ORDER BY embedding <-> '[3,1,2]' LIMIT 5;
embedding
------------------------------------
[0.9953359,0.916331,0.97390026]
[0.98834854,0.88819057,0.8903901]
[0.98210293,0.85718924,0.83997315]
[0.9161198,0.7464408,0.9253586]
[0.9961581,0.57789886,0.8373175]
(5 rows)
更多用法参考:
https://docs.vectorchord.ai/reference/ https://docs.vectorchord.ai/vectorchord/usage/indexing.html
参考
https://github.com/tensorchord/VectorChord/blob/main/sql/install/vchord--0.3.0.sql https://docs.vectorchord.ai/ https://blog.vectorchord.ai/vectorchord-store-400k-vectors-for-1-in-postgresql https://github.com/tensorchord/VectorChord https://github.com/tensorchord/VectorChord-bm25 https://github.com/tensorchord/pg_tokenizer.rs
VectorChord 的目标用户不局限于ANNs, 它还支持tokenize和rerank插件(下次分享), 云服务.