Vector databases

Pinecone vs pgvector

For most enterprises, this is the real vector DB choice: managed serverless or Postgres extension.

Quick verdict

pgvector for <10M vectors and existing Postgres ops. Pinecone for hands-off ops, elastic scale and multi-tenant serverless.

What each actually does

pgvector runs inside your Postgres. You get transactional consistency, ACL joins, backup and DR you already own. Simpler ops, harder scale ceilings.

Pinecone is a fully managed vector service. Elastic scale, sub-100ms latency at high QPS, no infra to run. You pay for that convenience.

The right answer depends on scale, ops appetite and how much your retrieval needs to join with relational data.

Pinecone

Fully managed serverless vector database with elastic scale and low-latency retrieval.

Strengths
  • Zero ops overhead
  • Elastic scale to billions of vectors
  • Consistent sub-100ms latency
  • Multi-tenant primitives built-in
Weaknesses
  • Pricing at scale gets expensive
  • Metadata filter perf can degrade at scale
  • Vendor lock-in (proprietary API)
  • No ACL joins with relational data

pgvector

Open-source Postgres extension for vector similarity search — runs where your Postgres does.

Strengths
  • Reuses existing Postgres ops
  • Transactional with relational data
  • ACL-aware retrieval via SQL joins
  • No new vendor, no lock-in
Weaknesses
  • Scaling beyond ~50M vectors gets hard
  • HNSW index build slow on big corpora
  • Requires DB tuning for high QPS
  • Multi-tenant patterns are DIY

Side-by-side

CriterionPineconepgvector
Ops overheadZero (managed)Same as your Postgres
Scale ceilingBillions of vectors~50M practical
ACL joins with dataExternalNative SQL
Cost at 10M vectors$$$$
Cost at 500M vectors$$Impractical
Lock-inProprietary APIOpen source, portable
Hybrid searchRequires external BM25SQL + pg_trgm / tsvector

Choose Pinecone when

  • Multi-tenant SaaS with elastic scale needs
  • Team doesn't want to run infra
  • Billions of vectors realistic in 12 months
  • Sub-100ms p95 latency non-negotiable

Choose pgvector when

  • Under 50M vectors realistic
  • Existing Postgres ops maturity
  • Retrieval must join relational data with ACLs
  • Cost sensitivity or open-source preference
The pragmatic answer

Use both, on purpose

Many teams start on pgvector, prove the app, then migrate hot indexes to Pinecone or a self-hosted alternative (Qdrant, Milvus) once scale demands it. Building against a portable abstraction (LangChain / LlamaIndex retriever) makes that migration a config change.

Frequently asked questions

What about Weaviate, Qdrant, Milvus?

All strong. Qdrant / Milvus win if you want open-source at big scale; Weaviate wins for built-in hybrid and modules. Compare them to Pinecone; pgvector remains the default first step.

How much scale can pgvector really handle?

Comfortably 10–20M vectors on a well-tuned instance. 50M+ needs careful HNSW tuning and read replicas. Beyond that, dedicated vector DBs pull ahead.

Can Pinecone do ACLs?

Via metadata filters and namespaces — workable but less expressive than SQL joins. For ACL-heavy enterprise, pgvector or Postgres-adjacent retrieval is often simpler.

Ready to Build Your AI Product?

Talk to a senior AI consultant from T7 about your industry, workflow, or product idea. Free, no commitment — reply within one business day.

  • · AI feasibility & architecture review
  • · Product / MVP roadmap
  • · Integration & automation strategy