Every RAG project eventually hits the vector database question. The marketing pages all sound the same. The real differences show up at scale, in production, when something breaks. Here's what we've learned from shipping on all of them.
The honest comparison
Pinecone | Weaviate | pgvectorPick | Qdrant | |
|---|---|---|---|---|
| Hosting model | Managed only | Managed + self-hosted | Self-hosted (Postgres) | Managed + self-hosted |
| Starting cost | $70/mo serverless | $25/mo managed | $0 (free with Postgres) | $0 (free self-hosted) |
| Cost at 10M vectors | $800+/mo | $400/mo | $50-100/mo | $100-200/mo |
| Query latency (p95) | < 50ms | < 80ms | < 100ms | < 30ms |
| Hybrid search (vector + keyword) | Basic | Excellent | Excellent (via extensions) | Good |
| Metadata filtering | ||||
| Learning curve | Easy | Medium | Easy (if you know SQL) | Easy |
| Best for | Quick managed start | Complex hybrid search | Already on Postgres | Self-hosted at scale |
When each one wins
Pinecone: the default for managed simplicity
Pinecone wins on developer experience. You sign up, get an API key, and you're indexing vectors in 10 minutes. Their serverless tier handles scaling automatically. For early-stage projects, prototypes, or teams that just don't want to manage infrastructure, Pinecone is the right answer.
Where it breaks: cost at scale. A few hundred thousand vectors is fine. Once you cross 5-10M vectors with production query volume, the monthly bill starts rivaling a small engineering salary. We've migrated multiple clients off Pinecone purely for cost reasons.
Weaviate: the hybrid search champion
Weaviate's killer feature is BM25 + vector hybrid search working cleanly out of the box. If your use case needs both keyword matching (for exact terms like product codes, names, IDs) AND semantic matching (for meaning), Weaviate is unusually good at blending them.
The tradeoff: more concepts to learn. Classes, cross-references, modules, and a custom query language (GraphQL). Worth the learning curve for complex search use cases, overkill for simple RAG.
pgvector: the underrated default
Here's the take most people don't expect: pgvector is our default for most new projects. Why? Because 90% of projects that need 'a vector database' already have Postgres for their application data. Adding pgvector is one extension install. Now your vectors live alongside your app data, with the same backups, same permissions, same connection pool, same SQL you already know.
Performance is 'good enough' for nearly everything — pgvector handles 10M+ vectors comfortably with proper indexing (HNSW). The 'specialized vector DB' advantage only matters above ~50M vectors or when query latency matters in the single-digit milliseconds.
“Don't add a new database unless your current one can't do the job. pgvector can do the job for 90% of RAG projects.”
Qdrant: the self-hosted scale winner
When you outgrow pgvector and don't want Pinecone's managed pricing, Qdrant is the answer. It's open-source, written in Rust, fast (fastest query latency of the four), and self-hostable on your own infrastructure for cents on the dollar compared to managed options.
The tradeoff: you're now operating a database. Backups, monitoring, upgrades — it's your responsibility. For teams with ops capacity, Qdrant is the best price/performance ratio. For teams without ops, stick with Pinecone or pgvector.
Our decision framework
Already using Postgres? → pgvector
Default answer. Unless you're doing something exotic, pgvector is cheapest, simplest, and good enough. Don't add infrastructure you don't need.
Enterprise project, managed only, budget not a concern? → Pinecone
Fastest way to get to production with zero infrastructure work. Expensive at scale but sometimes 'just handle it' is worth the money.
Complex search with hybrid keyword + vector? → Weaviate
If exact-match search matters alongside semantic search (legal docs, product catalogs, technical content), Weaviate's hybrid is worth the learning curve.
Self-hosted, 50M+ vectors, latency matters? → Qdrant
Best raw performance per dollar for teams that can manage their own infrastructure.
FAQ
Need help picking the right RAG stack?
We've built RAG systems on all four of these databases and migrated between them multiple times. If you're stuck on vendor choice or scale issues, book a 30-minute call and we'll give you an honest recommendation.
