Vector databases explained: pgvector, Qdrant, Pinecone and what to choose

A vector database is the heart of every RAG system. It lets you search documents by meaning, not just by words.

How it works

An embedding model turns each piece of text into a vector, for example a list of 1,024 or 1,536 numbers. The user's question is converted the same way. The database then finds the vectors "closest" to the question, which are the passages with the most similar meaning.

The most common choices

pgvector (PostgreSQL) An extension for PostgreSQL. Advantages: one database for all data, transactions, filtering by permissions and metadata with plain SQL. Our default choice.

Qdrant A dedicated open source vector database, excellent for large collections and advanced filtering. Can run on your own server.

Pinecone A fully managed cloud service. Fast to start with no maintenance, but data lives with a third party and cost grows with volume.

Weaviate and Chroma Weaviate for more complex schemas and hybrid search, Chroma for prototypes and smaller projects.

How we choose

• Up to a few million chunks, permissions required: pgvector.
• Tens of millions of vectors or very demanding search: Qdrant.
• Quick prototype without infrastructure: a managed service.
• Strict privacy requirements: a self-hosted solution on your infrastructure.

Hybrid search

The best results come from combining vector search with classic keyword search and re-ranking. This matters for article numbers, product codes and names that vector search can miss.

How much storage you need

Estimate size and cost with our RAG calculator. For 2,000 documents of 10 pages each, vectors usually take tens to a hundred megabytes, which is small.

Planning AI over your documents? Read what is RAG or see our AI document assistant service.