pgvector（pgvector 数据库）

pgvector Skill

PostgreSQL + pgvector extension for vector similarity search.

Quick Connect

# Connect to pgvector database (default port 5433)
psql -h localhost -p 5433 -U damien -d postgres

# Or use environment variables
export PGHOST=localhost
export PGPORT=5433
export PGUSER=damien
export PGPASSWORD=''
export PGDATABASE=postgres

Environment

• Host: localhost
• Port: 5433
• User: damien
• Password: (empty)
• Database: postgres

Core Capabilities

1. Create Vector Table

-- Basic vector table (1536 dimensions for OpenAI embeddings)
CREATE TABLE IF NOT EXISTS documents (
    id BIGSERIAL PRIMARY KEY,
    content TEXT NOT NULL,
    embedding vector(1536) NOT NULL,
    metadata JSONB,
    created_at TIMESTAMPTZ DEFAULT NOW()
);

-- Create HNSW index for fast similarity search
CREATE INDEX ON documents USING hnsw (embedding vector_cosine_ops)
WITH (m = 16, ef_construction = 64);

-- Or use IVFFlat index (faster build, slower search)
CREATE INDEX ON documents USING ivfflat (embedding vector_cosine_ops)
WITH (lists = 100);

2. Insert Embeddings

-- Manual insert (replace with actual embedding)
INSERT INTO documents (content, embedding)
VALUES ('Your text here', '[0.1, 0.2, ..., 0.1536]');

-- With metadata
INSERT INTO documents (content, embedding, metadata)
VALUES (
    'AI is transforming technology',
    '[0.1, 0.3, ..., 0.5]',
    '{"source": "article", "author": "John"}'::jsonb
);

3. Vector Similarity Search

-- Cosine similarity (most common)
SELECT id, content, (1 - (embedding <=> '[query_embedding]')) AS similarity
FROM documents
ORDER BY embedding <=> '[query_embedding]'
LIMIT 5;

-- Euclidean distance
SELECT id, content, (embedding <-> '[query_embedding]') AS distance
FROM documents
ORDER BY embedding <-> '[query_embedding]'
LIMIT 5;

-- Inner product (for normalized vectors)
SELECT id, content, (embedding <#> '[query_embedding]') AS similarity
FROM documents
ORDER BY embedding <#> '[query_embedding]'
LIMIT 5;

4. Hybrid Search (Vector + Keyword)

-- Combine vector search with full-text search
SELECT id, content,
    (1 - (embedding <=> '[query_embedding]')) AS vector_score,
    ts_rank(to_tsvector('english', content), plainto_tsquery('english', 'search terms')) AS text_score
FROM documents
WHERE content ILIKE '%search terms%'
ORDER BY (vector_score * 0.7 + text_score * 0.3) DESC
LIMIT 10;

5. RAG Pipeline Example

-- Store document chunks with embeddings
CREATE TABLE document_chunks (
    id BIGSERIAL PRIMARY KEY,
    document_id BIGINT REFERENCES documents(id),
    chunk_text TEXT NOT NULL,
    chunk_embedding vector(1536) NOT NULL,
    chunk_index INT NOT NULL
);

-- Retrieve relevant chunks for LLM context
SELECT chunk_text
FROM document_chunks
WHERE document_id = ?
ORDER BY chunk_embedding <=> '[question_embedding]'
LIMIT 5;

Management Commands

Check pgvector Extension

SELECT * FROM pg_extension WHERE extname = 'vector';

Table Info

-- List all tables with vectors
SELECT tablename FROM pg_tables WHERE schemaname = 'public';

-- Check index sizes
SELECT pg_size_pretty(pg_total_relation_size('documents'));

Monitoring

-- Check query performance
EXPLAIN ANALYZE
SELECT * FROM documents
ORDER BY embedding <=> '[query_embedding]'
LIMIT 5;

-- Index usage stats
SELECT * FROM pg_stat_user_indexes
WHERE indexname LIKE '%embedding%';

Common Operations

Update Embedding

UPDATE documents
SET embedding = '[new_embedding]'
WHERE id = 1;

Delete

DELETE FROM documents WHERE id = 1;

Batch Insert (Python)

import psycopg2
import numpy as np

conn = psycopg2.connect(
    host="localhost",
    port=5433,
    user="damien",
    password="",
    database="postgres"
)

cur = conn.cursor()
for text, embedding in documents:
    cur.execute(
        "INSERT INTO documents (content, embedding) VALUES (%s, %s)",
        (text, embedding.tolist())
    )
conn.commit()

Distance Operators

Use Cases

1. Semantic Search - Find documents by meaning, not keywords
2. RAG - Retrieve relevant context for LLM prompts
3. Recommendations - Find similar items/products
4. Anomaly Detection - Find outliers in embeddings
5. Image/Video Search - Store and query visual embeddings

Notes

• Vector dimensions must match your embedding model
• HNSW is better for accuracy, IVFFlat better for large datasets
• Normalize vectors for cosine similarity
• pgvector supports up to 16,000 dimensions