Vector database comparison posts in 2026 are mostly written by people building demo apps with 1,000 vectors. This is the version after running pgvector at production scale on Supabase, plus client work across Pinecone, Weaviate, and Qdrant. The actual choice in 2026 is less interesting than most people think — pgvector is the default for most projects, and the dedicated vector databases earn their place only at specific scales or specific feature requirements.
The five vector database options in 60 seconds
- pgvector (on Supabase, Neon, or your own Postgres) — Postgres extension that stores vectors alongside relational data. Free if you have Postgres. Right for 99% of projects under 10M vectors.
- Pinecone — managed vector database, mature platform, free tier 1 pod, Standard from $50/month, Enterprise custom. The default when 'I need a vector DB' is the explicit brief.
- Weaviate — open-source with Weaviate Cloud option, hybrid search built-in (vector + keyword), $25/month entry tier, scales to enterprise.
- Qdrant — open-source Rust-based, free Cloud tier 1GB, Hybrid Cloud and self-hosted options. Strong on performance per dollar at scale.
- Chroma — open-source, self-hosted-first, embedded mode for prototyping. Right for smaller projects and demos rather than production scale.
Where each option actually wins
pgvector: the default for 99% of projects
If you are already on Postgres — Supabase, Neon, RDS, self-hosted — pgvector is the answer for almost every vector workload under 10 million vectors. The vectors live alongside your relational data, you query them with normal SQL plus a `<=>` cosine distance operator, transactions work, RLS works for multi-tenant security. The HIPAA story works under the existing database BAA. The HIPAA-compliant Supabase setup covers pgvector in scope under the BAA.
- Wins on: integrated with relational data, free with existing Postgres, HIPAA in scope under database BAA, no extra vendor.
- Falls short on: scale past 10M vectors with high query throughput (specialized vector DBs perform better at that point), advanced hybrid search features that some specialized DBs include natively.
Pinecone: when the dedicated vector DB story is the brief
Pinecone is the dedicated vector database that most engineers think of first. The platform is mature, the documentation is good, the support is real, and the indexing performance at high scale is genuinely better than pgvector. Right call when you have a pure RAG application at scale (10M+ vectors, high QPS) or when the team is uncomfortable running Postgres extensions in production.
- Wins on: scale past 10M vectors, mature platform, dedicated vector workload optimisation.
- Falls short on: cost at small scale (Standard $50/month is more than most projects need), separation from relational data, the 'just one vendor' story.
Weaviate: hybrid search out of the box
Weaviate's differentiator is built-in hybrid search — combining vector similarity with keyword (BM25) ranking in one query. For RAG applications where the relevance tuning needs both semantic and keyword signals, Weaviate saves real engineering time. Open source with hosted Cloud option, scales to enterprise. Less polished than Pinecone but the hybrid feature is a real win for specific use cases.
- Wins on: hybrid search out of the box, open-source flexibility, hosted Cloud option.
- Falls short on: pure vector workloads where the hybrid feature is unused, mature platform polish versus Pinecone.
Qdrant: performance per dollar at scale
Qdrant is the Rust-based vector database with strong performance benchmarks per dollar. Open source, hosted Cloud free 1GB tier, Hybrid Cloud and self-hosted. Right call when you have a serious vector workload (10M+ vectors) and the cost-performance ratio matters more than the brand or the bundled features.
- Wins on: cost-performance at scale, open source, fast indexing.
- Falls short on: ecosystem maturity, AI personalization features, smaller community than Pinecone or Weaviate.
Chroma: embedded mode for prototypes
Chroma is the vector database for prototypes and small-scale projects. Embedded mode runs in-process; client-server mode is also supported. Right for development environments, demos, and projects under 1M vectors where the simplicity of 'pip install chromadb' beats spinning up a service. Less mature for production than Pinecone, Weaviate, or Qdrant.
- Wins on: prototyping speed, embedded mode for in-process use, simplicity.
- Falls short on: production scale, enterprise features, multi-tenant security.
Decision tree — pick by scale and stack
Already on Postgres, fewer than 10M vectors
pgvector. Free with your existing database, integrated with relational data, RLS works for multi-tenant. The default.
Pure RAG application at 10M+ vectors with high query throughput
Pinecone or Qdrant. Pinecone if mature platform and support matter; Qdrant if cost-performance matters and you are comfortable with the smaller ecosystem.
RAG with serious hybrid (semantic + keyword) requirements
Weaviate. The built-in hybrid search saves real tuning time.
Prototyping or development environment
Chroma in embedded mode. pip install, no service, get the demo working today. Migrate to pgvector or Pinecone when production calls for it.
FAQ
Is pgvector good enough for production RAG?
For most projects under 10M vectors, yes — pgvector at production scale on Supabase, Neon, or RDS handles the workload comfortably. The tipping point is roughly 10M vectors with consistent high QPS, where the dedicated vector databases start showing meaningful performance and cost advantages. Below that threshold, the operational simplicity of one database for everything beats the marginal performance gains of a separate vector service.
Is Pinecone HIPAA compliant?
Pinecone offers HIPAA on the Enterprise tier with a signed BAA. Below Enterprise, Pinecone is not HIPAA-eligible. If your project needs vector search under HIPAA, the simpler path is pgvector under your existing database BAA — Supabase Team with the HIPAA add-on covers pgvector in scope.
Should I use Chroma for production?
Generally no. Chroma is excellent for prototyping and development but lacks the multi-tenant security, replication, and production-grade ops that Pinecone, Weaviate, or Qdrant provide. Use Chroma to validate the RAG architecture, then migrate to a production-ready alternative.
Related reading
HIPAA-compliant AI in Next.js apps — vector databases in regulated industries, BAA scope context.
HIPAA-compliant Supabase + Vercel setup — pgvector under the Supabase BAA for healthcare RAG.
Serverless databases 2026 — when the database choice intersects with the vector search choice.
The vector database pick is rarely the AI app's bottleneck. The bottleneck is whether the embeddings, the chunking, and the retrieval pipeline are right. Pick the simplest store that lets you iterate on those.
Book a 30-minute vector / RAG call — describe the embedding model, the corpus size, the query pattern. Walk away with a pgvector-vs-Pinecone-vs-Qdrant decision that fits.