Picking a vector database in 2026 is more decision than benchmark. The five serious options are within a factor of two on raw performance for typical RAG workloads — the right pick depends on scale, deployment constraints, and team comfort. This piece is the broader survey across all five. For a tighter three-way head-to-head, see Pinecone vs Weaviate vs Qdrant in 2026.
What changed in 2026
- pgvector got fast. HNSW index in Postgres 17 is competitive with dedicated vector DBs up to mid-millions of vectors.
- Serverless became the default tier for managed vector DBs. Pay-per-query removes the capacity planning that was the operational pain of 2023.
- Hybrid search (BM25 + vector + reranker) is now table stakes — every serious option supports it. The differentiation is quality, not presence.
The five options
Pinecone. Closed-source, fully managed. Serverless-default in 2026. Easiest day-one experience; no self-host. Best when you want to ship today and don't want to think about infra.
Weaviate. OSS + managed. Strongest hybrid search story (modules, learned reranker). Both managed cloud and self-host are first-class. Best when retrieval relevance is the differentiator.
Qdrant. OSS + managed. Best price-performance at scale, especially with high-cardinality filters. Strong on-prem story. Best for cost- or residency-sensitive deployments.
pgvector. Postgres extension. Works with the database you already have. No additional service. Best for teams already on Postgres at small-to-mid scale (< ~5M vectors).
Milvus. OSS + managed (Zilliz). Built for very large deployments — billions of vectors. Strong distributed story. Best when you have very large corpora and a team that can run distributed infra.
Comparison table
| DB |
Self-host |
Hybrid search |
Best at scale |
Cost at 10M vectors |
| Pinecone |
No |
Good |
Yes (serverless) |
$$ |
| Weaviate |
Yes |
Excellent |
Yes |
$ |
| Qdrant |
Yes |
Good |
Yes |
$ |
| pgvector |
Yes |
Limited |
Up to ~5M |
$ (or free with PG) |
| Milvus |
Yes |
Good |
Very large |
$ |
How to choose by scale
- < 1M vectors, simple queries. pgvector. Don't add infra.
- 1M–10M vectors, simple to medium complexity. All five viable. Pick by deployment preference (managed → Pinecone, self-host → Qdrant or Weaviate).
- 10M–100M vectors. Pinecone serverless, Qdrant, Weaviate, or Milvus. Performance differences become real here; benchmark on your data.
- > 100M vectors. Milvus or Qdrant if self-host; Pinecone serverless if managed. Architecture matters more than tool branding.
How to choose by deployment shape
- Want zero ops and SaaS-only is fine. Pinecone.
- Need on-prem / VPC / strict residency. Qdrant or Weaviate.
- Already on Postgres and don't want a second service. pgvector.
- Already have a strong distributed-systems team and very large data. Milvus.
How to choose by retrieval quality
If retrieval relevance is the product (legal research, complex enterprise search), Weaviate's hybrid search ecosystem leads. Qdrant's sparse+dense fusion is competitive. Pinecone is good but less batteries-included. pgvector requires you to build your own hybrid pipeline.
What about benchmarks
Standard public benchmarks (BEIR, MTEB) tell you about embedding model quality more than vector DB quality. For DB choice, run a one-day benchmark on your own data and queries. Measure p50/p95/p99 latency at your target QPS, recall at your top-k, cost per query at expected volume. Two days of benchmarking saves a year of regret.
FAQ
Can I switch later?
Yes. Vector DBs are stateless from the LLM's perspective — re-embed and re-load. Plan for switching; don't fear it.
Is the embedding model more important than the DB?
For most applications, yes. A good embedding model with mediocre infrastructure beats the reverse.
What about Vespa or LanceDB?
Vespa is excellent for serious search teams; LanceDB is a strong newer option that's growing. Both worth evaluating; less common in typical 2026 RAG stacks.
Should I use a managed service or self-host?
Managed unless cost, residency, or compliance forces self-host. The TCO calc usually favors managed below ~50M vectors.
Where to go next
For related material see Pinecone vs Weaviate vs Qdrant in 2026, Best RAG tools for production in 2026, and RAG vs fine-tuning in 2026.