What is a Vector Database?
Traditional databases store strings and integers and search them using exact matches or wildcards. AI models, however, represent text as vectors—long arrays of numbers (e.g. 1536 dimensions) that represent the semantic meaning of sentences.
A vector database is optimized to store these numbers and execute "Cosine Similarity" or "Inner Product" calculations at scale to find which text snippets are closest in meaning to a user's question.
Why pgvector?
While specialized vector databases (like Pinecone, Milvus, or Qdrant) exist, pgvector is an extension for PostgreSQL that allows vector operations right alongside your core relational data. This offers immense benefits:
- No Extra Infrastructure: If you are already running PostgreSQL, you don't need to configure and pay for another standalone database.
- ACID Compliance: Get standard transactional guarantees on your vectors.
- Joint Queries: You can join vector search matches with traditional SQL tables (e.g.
WHERE org_id = X) in a single query.
How Similarity Search Queries Look
With pgvector, finding matching documents is as simple as running a SQL query:
SELECT id, content, embedding <=> :query_embedding AS distance
FROM documents
ORDER BY distance
LIMIT 5;
The <=> operator calculates cosine distance, sorting documents with the most similar concepts to the top. This simple extension enables fast, reliable semantic search architectures.