pgvector
Best when you already run PostgreSQL and your corpus is under ~10M vectors. Zero extra infra, ACID compliance, familiar SQL.
Pinecone
Managed, auto-scaling, metadata filtering. Good for teams that want to avoid ops overhead and can absorb the cost.
Weaviate
Open-source, hybrid search (vector + BM25) built in, strong GraphQL API. Our pick for self-hosted deployments that need flexibility.
Recommendation
Start with pgvector. Migrate only when you hit a clear scaling wall — most production RAG systems never need to.