How does it work?
An AI model turns text (or an image, or audio) into an embedding: a long list of numbers that captures its meaning. Two pieces of text that mean the same thing get vectors that sit close together, even if they use completely different words.
A vector store keeps all those vectors and can find the most similar ones in an instant. Picture a giant map where every document is a dot. Related topics sit close together. When someone asks a question, it becomes a dot on that same map, and the vector store returns the nearest neighbours.
A traditional database searches for exact words. A vector store searches for meaning.
What is it for?
Whenever you want to search information by meaning, a vector store comes into play. The most common uses:
- Semantic search: searching for what someone means, not the exact term. Someone searching for "lower energy costs" also finds a page about "saving on heating bills".
- AI assistants on your own data (RAG): the assistant pulls the right passages from your content and uses them to form an answer, with a source reference.
- Recommendations: suggesting similar products, articles or files.
- Classification and deduplication: automatically recognising what belongs together or is a duplicate.
Why do you really need it?
An ordinary search function works on keywords. If the exact word isn't there, it finds nothing, even if the text is a perfect match in substance. For a ten-question FAQ that's no problem. For an AI assistant that has to answer reliably across hundreds of pages of content, emails or documents, it is.
The vector store is exactly the piece that makes sure the assistant gets the right information before it answers. Without that layer a language model guesses, or makes something up. With it, the model answers based on your real content, and can point to the source.
Vector store or vector database?
The terms are used interchangeably and in practice mean the same thing: a system that stores vectors and makes them searchable. Sometimes it's a separate, specialised database, sometimes an extension to an existing one. We often use pgvector, an extension to PostgreSQL, so your vectors live next to your regular data in one system you already know.
Related terms
- Embeddings: the number sequences that capture meaning and end up in the vector store.
- RAG (retrieval-augmented generation): the technique where an AI first retrieves relevant info from the vector store and only then answers.
- Semantic search: searching by meaning, made possible by embeddings and a vector store.
