Vector stores: where meaning gets indexed
What it is — in one coffee-break
A vector store is the database where embeddings live (lesson 2 covered what those are: meaning, expressed as numbers). Its one core trick is nearest-neighbour search: given the embedding of a question, return the stored texts whose embeddings sit closest — in milliseconds, even across millions of entries. Checking every stored vector one by one stops working at around ten thousand entries, so vector databases use clever index structures (HNSW is the default everyone uses) that trade a sliver of accuracy for thousand-fold speed.
When you actually need it (and when you don't)
You need one exactly when you need embeddings at scale: a knowledge assistant over thousands of documents, semantic search across your ticket history, deduplication across a big catalogue. You don't need a separate one nearly as often as vendors suggest. The 2026 practitioner consensus is refreshingly boring: if you already run Postgres — and most businesses do, often without knowing it — the free pgvector extension handles up to ~10 million vectors alongside your normal data, with one backup story and zero new infrastructure. Our SOP answer machine and receipt archive blueprints follow the same keep-it-simple logic.
Dedicated stores (Qdrant, Weaviate — both open-source and self-hostable — or managed Pinecone) earn their place at tens of millions of vectors, or when you need specific features like multi-tenancy. And whatever the store, hybrid search is the production standard: pure meaning-based search misses exact terms like error codes and product SKUs, so real systems combine it with classic keyword search.
How to recognize good vs bad implementations
In the audits, the Vector store element (Vx) again exposes platform strategy. Gemini scores 9 — Google sells the whole retrieval stack, storage included. ChatGPT scores 7 with built-in file search that just works at small scale. Claude scores 2 — bring your own, same deliberate gap as embeddings. Workflow tools cluster mid-table (n8n 6.5, Make 5.5) with connectors to the usual stores. The trap to recognize: tools that store your embeddings but won't let you export them. That's your data, indexed — if it can't leave, the meaning-layer of your knowledge base is locked in.
What this costs
Storage is the meter, and dimensions drive it: a million vectors at 1,536 dimensions is roughly 6 GB before index overhead. Self-hosted pgvector: effectively €0 plus the server you already have. Managed serverless starts around $0.33/GB-month — trivial at small business scale, meaningful at millions of documents. The cost levers are dimension truncation and quantization: shrinking vectors 6–12× costs only a few percent of quality. For almost everyone reading this: your vector bill will be under €10/month, and pgvector makes it zero.
Where to see it scored
Vx with reasons: Gemini (9), ChatGPT (7), Windsurf (6.5), Claude (2 — bring your own). Next lesson: RAG — the pattern that puts embeddings, the vector store and the model together.