Supabase pgvector: Building Vector Search Without the Complexity

The Case for Unified Data

The default advice for AI applications is to use a dedicated vector database like Pinecone, Weaviate, or Qdrant. For many teams, this is overkill. If your data already lives in Postgres, adding pgvector to your existing Supabase instance eliminates an entire infrastructure component.

Setting Up pgvector in Supabase

Supabase ships with pgvector enabled by default. Creating a vector column is as simple as:

ALTER TABLE documents ADD COLUMN embedding vector(1536);

For similarity search, create an index:

CREATE INDEX ON documents USING ivfflat (embedding vector_cosine_ops) WITH (lists = 100);

Performance Reality

We have tested pgvector extensively and here are the honest numbers:

• Under 100K vectors: pgvector matches or beats dedicated vector databases in query latency
• 100K-1M vectors: pgvector is competitive with proper indexing (HNSW recommended over IVFFlat)
• Over 1M vectors: Dedicated vector databases start showing advantages in query performance and scaling

For most startups and mid-stage products, you will never hit 1M vectors. Your product documentation, customer support tickets, and knowledge base articles probably total 10K-50K documents.

The Killer Feature: Hybrid Queries

The real advantage of pgvector in Supabase is joining vector similarity with relational filters in a single query. "Find the most similar documents to this query, but only from this customer's workspace, created in the last 30 days, with status = published." Try doing that efficiently with a separate vector database.

When to Use a Dedicated Vector DB

If you have more than 1M vectors, need sub-10ms query latency, or require advanced features like namespaces and metadata filtering at scale, a dedicated solution makes sense.