PGVector Tutorial 2026: Add AI Search to PostgreSQL in 30 Minutes

You already use PostgreSQL - adding vector search keeps your stack simple. PGVector is the most popular way (22,400 stars).

## PGVector Tutorial 2026: Add AI Search to PostgreSQL in 30 Minutes PGVector is the PostgreSQL extension that adds vector storage and similarity search to your existing database. At 22,400 GitHub stars, it is the most popular way to add semantic search without introducing a separate vector database. If you already run Postgres, this tutorial gets you a working RAG search in half an hour. ## Why PGVector Instead of a New Database | Factor | PGVector | Dedicated vector DB | |--------|----------|---------------------| | Setup | 1 command | New service to run | | Data consistency | Same DB as your data | Sync needed | | Transactions | Full ACID | Varies | | Scale | Great to ~1M vectors | Better at 100M+ | | You already run Postgres | Free | Extra cost | ## Step 1: Install the Extension ```sql CREATE EXTENSION IF NOT EXISTS vector; ``` On Ubuntu: apt install postgresql-16-pgvector. On Mac: brew install pgvector. ## Step 2: Create a Table ```sql CREATE TABLE documents ( id SERIAL PRIMARY KEY, content TEXT, embedding vector(768) ); ``` The 768 dimension must match your embedding model (bge-m3 uses 1024; nomic-embed-text uses 768). ## Step 3: Insert Embeddings (Python) ```python import psycopg2 from openai import OpenAI client = OpenAI(base_url="http://localhost:11434/v1", api_key="ollama") vec = client.embeddings.create(model="nomic-embed-text", input="your text").data[0].embedding conn = psycopg2.connect("dbname=app") conn.execute( "INSERT INTO documents (content, embedding) VALUES (%s, %s)", ("your text", vec) ) ``` ## Step 4: Search ```sql SELECT content, 1 - (embedding <=> query_embedding) AS similarity FROM documents ORDER BY embedding <=> query_embedding LIMIT 5; ``` The <=> operator is cosine distance. Lower = more similar. ## Step 5: Add an Index (required at scale) ```sql CREATE INDEX ON documents USING hnsw (embedding vector_cosine_ops); ``` HNSW indexes make search milliseconds instead of seconds once you pass ~10K rows. ## FAQ **PGVector vs dedicated vector DBs?** PGVector wins on simplicity and consistency for most apps under ~10M vectors. Chroma (28,900 stars), Qdrant (33,700), and Milvus (45,400) win at massive scale or when you want a purpose-built store. **Does it work with LangChain?** Yes - langchain-postgres has a PGVector vectorstore class. **HNSW or IVFFlat index?** HNSW - better accuracy and no tuning needed. IVFFlat is legacy.