What Is pgvector? A Practical Introduction

pgvector is an open-source extension that adds vector similarity search to PostgreSQL. In plain terms: it lets your database store the numeric "embeddings" that AI models produce and find the ones most similar to a given query — turning ordinary PostgreSQL into a capable vector database, with no separate system to run.

Why vectors matter

Modern AI models turn text, images, or audio into a list of numbers called an embedding — typically a few hundred to a couple thousand floating-point values. Similar items produce nearby vectors. "Find similar" becomes "find the nearest vectors," which is the foundation of semantic search, recommendations, and retrieval-augmented generation (RAG). pgvector gives PostgreSQL a native way to store and search them.

What pgvector adds

Three things:

1. A vector column type — store an embedding as a first-class column, e.g. embedding vector(1536).
2. Distance operators — <-> (Euclidean/L2), <=> (cosine), and <#> (negative inner product) measure how close two vectors are.
3. Approximate-nearest-neighbour indexes — HNSW and IVFFlat make search fast at scale instead of scanning every row.

Everything else is just PostgreSQL: transactions, joins, WHERE filters, backups, replication.

A first example

CREATE EXTENSION vector;                 -- enable it (a no-op on Rivestack)

CREATE TABLE items (
  id        bigserial PRIMARY KEY,
  content   text,
  embedding vector(1536)
);

-- after inserting rows with embeddings:
CREATE INDEX ON items USING hnsw (embedding vector_cosine_ops);

-- find the 5 most similar items to a query vector:
SELECT content
FROM items
ORDER BY embedding <=> $1
LIMIT 5;

That ORDER BY embedding <=> $1 is the whole idea — rank rows by cosine distance to a query embedding, fast, using the HNSW index.

What pgvector is used for

• RAG — retrieve the most relevant chunks to ground an LLM's answer.
• Semantic search — search by meaning, not keywords.
• Recommendations — "more like this" over products, articles, or media.
• Deduplication and clustering — group near-identical content.

Because the vectors sit next to your relational data, you combine similarity with normal filters in one query — see building RAG on PostgreSQL.

How to start

Any PostgreSQL with the extension works. On a managed service it is usually available to enable; on Rivestack pgvector 0.8.x ships pre-installed and tuned, so CREATE EXTENSION vector is a no-op and you can insert embeddings immediately. From there, the getting-started guide walks through your first index and query.

FAQ

What is pgvector used for?

Storing embeddings and finding similar ones: retrieval-augmented generation (RAG), semantic search, recommendations, and deduplication. Anywhere you have AI-model embeddings and want nearest-neighbour search, optionally combined with SQL filters.

Is pgvector free?

Yes. pgvector is open-source (PostgreSQL license). You only pay for the PostgreSQL instance you run it on — and Rivestack's free tier includes pgvector at no cost.

What's the difference between HNSW and IVFFlat?

They are the two index types pgvector offers. HNSW is a graph index with the best query speed and recall (the production default); IVFFlat uses inverted lists and is cheaper to build but usually needs more tuning. See our HNSW vs IVFFlat guide.

---

Learn more on the pgvector guide, or read about PostgreSQL for AI.