The right vector database depends on your scale, query patterns, filtering requirements, and whether you want managed vs. self-hosted infrastructure. Here's what the choice actually looks like in 2026.
2026 comparison table
| Database | Managed | ANN QPS (1M vecs) | Metadata filtering | Cost at 100M vecs | Best for |
|---|---|---|---|---|---|
| Pinecone | Yes | ~2,000 | Strong | ~$2,000/mo | Managed, serverless, enterprise |
| Weaviate | Cloud + self | ~1,800 | Very strong | ~$800/mo self | Multi-modal, complex filtering |
| Qdrant | Cloud + self | ~3,500 | Strong | ~$400/mo self | High QPS, cost efficiency |
| pgvector | Any Postgres | ~200 | Full SQL | Near zero | <1M vecs, existing Postgres stack |
| Chroma | Self-hosted | ~300 | Basic | Free | Prototyping only |
How to pick
- Under 1M vectors, existing Postgres: pgvector. Zero infrastructure overhead, full SQL filtering, adequate performance at this scale.
- 1M–50M vectors, need managed: Pinecone. Serverless tier scales smoothly; operational simplicity worth the premium for most teams.
- High QPS, cost-conscious, okay with ops: Qdrant. Best raw performance per dollar — HNSW + quantization.
- Complex metadata + multi-modal: Weaviate. Its schema system handles complex filtered query patterns better than alternatives.
- Prototyping only: Chroma. Don't run it in production.
The filtering trap
Most vector databases degrade badly when you add metadata filters — they fall back from ANN to brute-force on the filtered subset. If your queries require filtering on low-cardinality fields (tenant_id, category), test filtered query performance specifically. Qdrant and Weaviate handle this best.
Any vector database on MoltBot
Pinecone, Weaviate, Qdrant, and pgvector all natively supported. 14-day free trial.
Start Free Trial →