Picking a vector DB in 2026: pgvector, Pinecone, Weaviate

On most new RAG projects this is the first real architecture decision: where does the vector DB live. Get it wrong and you're not just out some money — you're out months migrating later. Three serious contenders: pgvector, Pinecone, Weaviate.

pgvector — the 'just Postgres' route

If you already run Postgres, pgvector is usually the right first pick. Extension, no new system, transactional data and embeddings in the same DB. Cost: effectively zero. Performance: comfortable up to ~1M vectors. IVFFlat + HNSW supported.

• Pros: zero new infra, JOINs work, row-level security works.
• Cons: tuning IVFFlat past 10M vectors is non-trivial.
• Sweet spot: 0 to ~1M vectors, per-customer knowledge bases.

Pinecone — managed and fast

If you don't want to run a vector index and want an API call away, Pinecone makes sense. Multi-tenant serverless, metadata filtering, autoscale. Cost: scales up faster than you expected if sharding is wrong.

• Pros: zero ops, fast ramp, stable tail latency.
• Cons: vendor lock, no JOIN back to transactional data, expensive past 10M vectors.
• Sweet spot: fast MVP, SaaS scale-outs across thousands of tenants.

Weaviate — the hybrid take

Weaviate plays a different game: native hybrid search (BM25 + vector + filter), GraphQL API, modules (reranker, generator). Self-host or managed. Stronger for document-centric use cases.

• Pros: native hybrid search, rich schema, multi-tenant.
• Cons: steeper learning curve; ops overhead if self-hosted.
• Sweet spot: knowledge-base search, document platforms, hybrid-first retrieval.

Our decision framework

1. Already have Postgres and < 1M vectors? → pgvector.
2. Multi-tenant SaaS, < 10M per tenant, fast launch? → Pinecone.
3. Large document corpus, hybrid search is first-class? → Weaviate.
4. Unsure? Start with pgvector — the switching cost is lower than vendor lock-in.

Takeaway

There's no universal answer. There is a framework. Happy to look at your setup — 30 minutes is enough for a directional recommendation.