Vector Databases: When You Need One and When You Don't

pgvector, Pinecone, or just an in-memory index? A practical guide to choosing storage for your AI features.

The vector database hype made it feel mandatory, but most apps reach for one too early. I break down the real decision: at small scale, pgvector inside the Postgres you already run beats adding a new service. At larger scale, dedicated vector stores earn their keep with filtering, sharding, and managed scaling. I benchmark recall versus latency trade-offs, explain HNSW versus IVF indexes in plain terms, and share a rule of thumb for when to graduate from one tier to the next—so you spend complexity budget only where it buys you something.