Pick a vector database
TinySet compares vector databases on license, hosting and price. TinyResearch weighs fit and lock-in for a large RAG app.
TinySet
Dataset · 49 rowsVector databases, with license, hosting options, starting price and max dimensions
| Vector database | License | Hosting options | Starting price |
|---|---|---|---|
| Pinecone | Proprietary | cloud | $0 free tier |
| Weaviate | BSD-3-Clause | cloud, self-hosted | $0 free tier |
| Milvus | Apache-2.0 | self-hosted | $0 free tier |
| Qdrant | Apache-2.0 | cloud, self-hosted, hybrid | $0 free tier |
| Chroma | Apache-2.0 | cloud, self-hosted | $0 free tier |
| Faiss | MIT | self-hosted | |
| pgvector | PostgreSQL | self-hosted | $0 free tier |
| Elasticsearch | Elastic-2.0 | cloud, self-hosted |
TinyResearch
ReportWhich vector database fits a 10M-document RAG app, and what are the lock-in risks?
A 10-million-document RAG application sits at a critical inflection point in the vector database landscape: it is too large for pgvector alone without careful engineering, comfortably within the operational range of dedicated engines (Qdrant, Weaviate, Pinecone), and well below the threshold where Milvus's operational complexity is truly justified. The decision turns less on raw latency—most mature engines achieve sub-50ms at this scale—and more on three overlapping factors: hybrid search quality (now a production requirement, not a differentiator [3][47]), total cost of ownership including hidden billing mechanics, and lock-in exposure.
