Vector Databases in 2025: Which to Pick When
Pinecone vs Weaviate vs Qdrant vs pgvector vs Chroma — the honest comparison.
"Which vector database should I use?" is the most common question we get from technical founders. The honest answer is "almost always pgvector," but here's the full decision tree.
By scale
- Under 1M vectors: pgvector. You probably already have Postgres. Adding pgvector is one extension. No new service to operate, no new query language, transactions and joins work normally.
- 1M–10M vectors with multi-tenancy: Qdrant self-hosted. Best balance of performance, operational simplicity, and metadata filtering at this range.
- Above 10M vectors, multi-region, zero-ops: Pinecone. You're paying for managed scale; that's exactly what it's good at.
- Local dev and prototyping: Chroma. Embedded, no infrastructure, two lines to install.
By feature requirement
- Hybrid search (keyword + vector): Weaviate, Qdrant, or pgvector + Postgres full-text. Pinecone's hybrid is catching up but still less mature.
- Self-hosted with enterprise support: Weaviate or Qdrant. Both offer commercial editions.
- Sharded multi-tenant at SaaS scale: Pinecone or Turbopuffer.
- Tight Postgres integration (joins, transactions): pgvector, no contest.
The mistake that costs the most money
Jumping to managed Pinecone at 50K vectors because "we'll scale eventually." You're paying enterprise prices for hobby-scale data, and pgvector would have been faster anyway because the network hop dominates. Start with pgvector; move when the data forces you to.
The mistake that costs the most engineering time
Picking a vector DB before you've designed your metadata schema. You'll always end up filtering by tenant_id, document_type, date range, language — and the DB you picked may make that fast, slow, or impossible. Sketch the queries first, pick the DB second.
What none of them solve well
- Re-embedding migrations at scale — that's your problem, not the DB's.
- Embedding model versioning — track it in metadata yourself.
- Sparse + dense hybrid scoring that's actually tunable — most "hybrid" implementations are first-pass.
Bottom line
Default to pgvector. Switch to Qdrant when you outgrow it. Switch to Pinecone when you outgrow Qdrant and don't want to operate infrastructure. Don't pay for scale you haven't reached.