What Is a Vector Database

If your agent needs to know things — your refund policy, your product specs, your customer's history — a vector database is how it accesses that knowledge. Traditional databases require exact queries: SELECT * WHERE category = 'billing.' Vector databases handle fuzzy, natural language queries: find documents related to 'the customer wants their money back.'

I use Supabase with pgvector for most client projects because it eliminates the need for a separate service. The business data and the vector store live in the same database, which simplifies architecture, reduces costs, and makes the deployment easier to manage. For larger-scale deployments where you're storing millions of documents, Pinecone or Weaviate handle the volume better.

The setup process is straightforward: take your documents (knowledge base articles, product pages, FAQs, SOPs), split them into chunks, generate embeddings using an embedding model (OpenAI's text-embedding-3-small works well), and store the chunks with their embeddings in the vector database. When the agent needs information, it converts its query to an embedding and searches for the closest matches.

Traditional databases search by exact matches — find records where the email equals 'john@example.com.' Vector databases search by meaning — find documents that are similar to 'What's your refund policy for annual subscriptions?' The query doesn't need to match any stored document word-for-word. It finds documents that are semantically close.

This works through embeddings: numerical representations of text that capture meaning. 'What's your refund policy?' and 'How do I get a refund?' produce similar embeddings even though they use different words. The vector database stores these embeddings and uses algorithms (HNSW, IVF) to find the closest matches quickly — even across millions of stored vectors.

The most common pattern is RAG: store your business knowledge as vector embeddings, and retrieve relevant chunks when the agent needs information. Support agents search the knowledge base for answers. Sales agents search product documentation for specs. Operations agents search SOPs for procedure details.

Beyond RAG, vector databases power agent memory. Store summaries of past interactions as vectors, and the agent can retrieve relevant history when a customer returns. 'Last time this customer contacted us, they were frustrated about delivery delays — approach with empathy.' That contextual memory transforms the agent's effectiveness.