Vector databases in 2025: pgvector, Pinecone, Weaviate, Qdrant
Honest trade-offs after running production RAG on each one.
If your data already lives in Postgres, pgvector is the right starting point in 2025. It's no longer a toy: HNSW indexes, halfvec types and filter pushdown make it competitive with dedicated vector stores for workloads under ten million vectors. You also get joins, RLS and transactions for free.
Pinecone shines when you need fully managed, multi-region, low-latency reads at scale and you don't want to operate anything. The price is real and the lock-in is real, but for teams that value zero-ops it's the cleanest path.
Qdrant and Weaviate are both excellent open-source options with rich filtering, hybrid search and decent self-hosted stories. Choose Qdrant for raw performance and a simple mental model, Weaviate when you want the schema-rich, GraphQL-flavored developer experience.