Four vector databases compared — Pinecone, Qdrant, Weaviate, Chroma
Which retrieval layer to put under a RAG system. Prices were checked on each vendor's pricing page in November 2026.
AI Tools Hub Editorial · published · updated
At a glance
| Pinecone | Qdrant | Weaviate | Chroma | |
|---|---|---|---|---|
| Free tier | 5 indexes, 2GB, one region | 0.5 vCPU, 1GB, forever | 100k objects, 1GB | $0 plus $5 credits |
| Entry paid | $20/mo flat | Usage-based | From $45/mo | $250/mo (Team) |
| Open source | No | Yes | Yes | Yes |
| Run it yourself | No | Hybrid / Private Cloud | Enterprise Edition | Self-host the OSS |
| Embeddings included | Yes | Partly on free tier | Yes | No |
| Full-text search | — | — | — | Same engine |
By scale
Prototype, under 100k vectors
All four fit in the free tier. Pick whichever client library you find pleasant and move on; there is nothing to optimise yet.
Small production, under one million
Pinecone's flat $20 Builder tier generally covers this, and not having to think about usage is worth the money at this stage.
Larger
Usage charges start to matter and self-hosting becomes arguable. Remember that self-hosting costs server time plus a person's attention for updates, backups and incidents. Saving a few hundred dollars a month while tying up an engineer is a bad trade.
When you also need exact matching
Vector search is weak on part numbers and exact identifiers. A query about product code XYZ-100 wants lexical matching. Chroma handles vector, full-text and metadata in one engine, which saves running a second search system. The others can be paired with Elasticsearch to the same effect, at the cost of another component.
Three questions that settle it
- Can the data leave your infrastructure? If not, one of the open-source three.
- Do you want embeddings from the same vendor? If so, Pinecone or Weaviate.
- Do you need exact matching too? If so, Chroma.
FAQ
What about pgvector?
If you already run PostgreSQL and have a few hundred thousand vectors, it is a sound choice. Dedicated databases earn their place at larger scale or with complex filtered queries. Not adding another database to operate has real value.