pgvectorvsPinecone
TL;DR
pgvector lives next to your business data and wins for almost every workload up to 100M vectors. Pinecone earns its keep for billion-scale or when you want zero database operations. For most teams, the bundled-with-Postgres model is the calmer pick.
Sub-100M vectors
pgvector
Billion-scale, big budget
Pinecone
Joins with business data
pgvector
Zero database ops
Pinecone
| Feature | pgvector | Pinecone |
|---|---|---|
| Hosting model | Extension in your Postgres | Managed SaaS only |
| Index type | HNSW (default), IVFFlat | HNSW-based proprietary |
| Hybrid search | RRF over vector + lexical SQL | Built-in sparse + dense |
| Joins with tables | First-class (same SQL) | Not possible (separate system) |
| Authz / RLS | Postgres RLS works | API-key + namespace separation |
| Pricing | Existing Postgres compute | Per-vector / per-query |
| Quantisation | PQ in 0.7+, halfvec / bit | Proprietary; on by default |
| Scale ceiling | Tens to hundreds of millions | Billions |
| Operational footprint | Same as Postgres | Zero, fully managed |
When pgvector wins
- Your vectors are part of a larger application that already lives in Postgres. Joining a top-K vector search with a users table or a tenant filter is one SQL query.
- You want RLS to apply to your vector search the same way it applies to the rest of your data.
- Your scale is under ~100M vectors. HNSW on a well-sized Postgres handles this with latency comparable to Pinecone's serverless tier.
- You don't want a second cloud bill or a second monitoring story.
When Pinecone wins
- You're at billion-vector scale. Pinecone's proprietary indexing and multi-region serverless were built for this.
- You don't want to operate a database at all. Pinecone is the most hands-off of the vector stores.
- You have an enterprise procurement process that prefers a dedicated SaaS vendor.
Honest take
The 2026 default for production RAG is pgvector. It hit feature parity with the dedicated vector stores in 2024, and the join-with-business-data advantage compounds as your application grows. Pinecone is still a great product; it's sized for a workload (billions of vectors, zero database ops, enterprise pricing) that most teams don't actually have. If you do have that workload, Pinecone is the safer pick; otherwise, you're paying for capacity you won't use.