Qdrant

Qdrant is an open-source vector database and similarity-search engine written in Rust, used as the retrieval layer for AI applications. It stores vectors alongside JSON metadata and serves nearest-neighbour queries over an HNSW index, with filters applied during index traversal in a single stage rather than before or after the search, so filtered queries keep recall high without a separate scan.

Search is hybrid by default: dense and sparse vectors can be combined in one query, with support for BM25, SPLADE++, and miniCOIL. Multiple vectors per object allow multimodal and multi-representation retrieval, and results can be shaped further with score boosting for business logic, late-interaction models such as ColBERT for token-level precision, and Maximum Marginal Relevance for diversity. Metadata filtering covers nested fields, text, geographic conditions, and vector presence.

Deployment spans several models from the same engine: a fully managed cloud with high availability and auto-sharding on AWS, Google Cloud, or Azure; a hybrid cloud that separates control and data planes so the data plane runs in your own Kubernetes; a private cloud for air-gapped and compliance-bound installations; and a beta edge build for low-latency search near where data is produced. Operational features include multitenancy with granular role-based access control, backups with point-in-time restore, and metrics for Prometheus, Grafana, and Datadog.

Features

• Vector search engine: HNSW-based approximate nearest-neighbour search written in Rust
• One-stage filtering: filters are applied during index traversal, avoiding pre- or post-filtering trade-offs
• Hybrid search: dense and sparse vectors in a single query, with BM25, SPLADE++, and miniCOIL support
• Multiple vectors per object: several representations per record for multimodal and multi-view retrieval
• Relevance controls: score boosting, late-interaction models such as ColBERT, and Maximum Marginal Relevance
• Metadata filters: JSON payloads with nested, text, geographic, and vector-presence conditions
• Deployment models: managed cloud, hybrid cloud on your own Kubernetes, private air-gapped cloud, and an edge build
• Operations: multitenancy and granular RBAC, backups with point-in-time restore, and Prometheus, Grafana, and Datadog metrics