The rise of embeddings-based search means most teams shipping an AI feature now need a vector database somewhere in the stack — and the open-source options have matured fast.
Vector databases
Qdrant and Weaviate are both open-source, self-hostable, and offer a managed cloud tier if you'd rather not run the infrastructure yourself. Pinecone takes the opposite bet: managed-only, no self-hosted option, in exchange for not having to think about the underlying cluster at all.
None of these is strictly "better" — the right pick depends on whether you want to own your infrastructure, and how tightly you need dense, sparse, and full-text search to sit next to each other in one query.
Relational holdouts
Postgres remains the default for a reason: pgvector now lets a lot of teams skip a dedicated vector database entirely for small-to-medium corpora, at the cost of the specialized indexing and query performance a purpose-built vector store offers at scale.
The actual decision
Start with what you already run in production. Adding a new database is a real operational cost — the calculus only shifts once your existing stack's vector support is measurably too slow for what you're building.