qdrant

It stores vectors alongside their associated payload metadata, enabling hybrid queries that combine semantic similarity search with structured metadata filteringfinding semantically similar items that also match specific attribute criteria.

Qdrant is implemented in Rust for performance and memory efficiency, and is designed for both self-hosted deployment and Qdrant Cloud managed service.

The database implements HNSW (Hierarchical Navigable Small World) indexing with several production-focused extensions: scalar and product quantization to reduce memory requirements for large collections, on-disk vector storage for collections that exceed GPU/CPU RAM, payload indexing for efficient metadata filtering, and sparse vector support for hybrid dense+sparse retrieval strategies.

Qdrant's filtering during HNSW searchrather than post-filtering after retrievalmaintains high recall accuracy even with restrictive metadata filters that would otherwise reduce effective recall significantly.

Teams building semantic search, recommendation systems, RAG pipelines, image similarity search, and anomaly detection applications use Qdrant as their vector storage layer.

Its combination of developer-friendly REST and gRPC APIs, well-documented Python client, and performance characteristics competitive with or exceeding dedicated cloud vector services has driven strong adoption in the AI application development community.

The Rust implementation's reliability and memory safety properties make it a compelling choice for production deployments where stability matters, and its Docker-based self-hosting option provides full control over data residency for privacy-sensitive applications.