Qdrant

If you're picking a vector database, the default answer most teams reach for is Pinecone: managed, closed and quick to start. Qdrant makes a different bet. It's an open-source engine written in Rust, with 35k+ GitHub stars and 60k+ community members, and you can run the same engine in several places: self-hosted, on Qdrant Cloud across AWS, GCP or Azure, on your own Kubernetes through Hybrid Cloud, air-gapped in Private Cloud, or close to where data is generated with Qdrant Edge (Beta). That deployment range is the main reason to choose it over a managed-only service. You aren't tied to one vendor's hosting to get the query engine you built against.

The engine itself is where Qdrant earns its reputation. Filtering happens during HNSW traversal instead of as a separate pre- or post-filter step, which Qdrant says keeps recall high and latency low even under complex conditions. Metadata is stored as JSON and supports nested, text, geo and has_vector filters. Hybrid search that blends dense and sparse vectors is native, with BM25, SPLADE++ and miniCOIL support. You also get multiple vectors per object, plus reranking options like score boosting, ColBERT-style late interaction and Maximum Marginal Relevance. That's a lot of retrieval tuning living in the database layer rather than scattered through your app code.

The tradeoffs are worth knowing up front. Qdrant Serverless is still marked "coming soon," Edge is in beta, and the pricing page we pulled didn't show tier details beyond a free starting point, so check qdrant.tech before you budget. On the enterprise side, the checklist is solid: SOC 2 and HIPAA compliance, SSO via SAML or OIDC, multitenancy with granular RBAC, private networking, zero-downtime upgrades, point-in-time restore and monitoring through Prometheus, Grafana or Datadog. If you want a vector store that can go from a laptop prototype to a production RAG or agent system without swapping engines, Qdrant is one of the strongest picks available, provided you're comfortable owning a bit more infrastructure thinking than a pure serverless product asks of you.