A Rust vector database that treats embeddings like database rows with JSON metadata

What it does

Qdrant is a vector database and similarity search engine written in Rust. It stores “points”—vectors with attached JSON payloads—and provides APIs to search, filter, and manage them. The payload system is the hook: you can attach arbitrary metadata and filter on keywords, ranges, geo-locations, and booleans using should, must, and must_not clauses, then combine those filters with vector similarity search.

The interesting bit

Most vector stores treat metadata as an afterthought; Qdrant built payload indexing into the query planner from the start. It also supports dense, sparse, and multivector search in the same query—meaning you can fuse semantic embeddings, keyword sparse vectors, and late-interaction models like ColBERT, then merge results with RRF or DBSF. The quantization claims up to 97% RAM reduction, which is a concrete number worth testing.

Key highlights

• Hybrid search fuses multiple vector types with configurable score fusion strategies
• Built-in quantization and on-disk storage for memory-constrained deployments
• Distributed mode with sharding, replication, and zero-downtime collection resizing
• GPU-accelerated indexing for NVIDIA and AMD cards
• io_uring async I/O for disk throughput on network-attached storage
• Edge variant (Qdrant Edge) runs embedded in application processes for offline/low-latency use
• Web UI included for visual collection management and API exploration

Caveats

• The README’s “97% RAM reduction” claim lacks independent verification in the source; treat as vendor benchmark
• Default Docker deployment is intentionally insecure (no auth, all interfaces open), which is noted but easy to miss

Verdict

Worth evaluating if you’re building semantic search, recommendation, or RAG systems and need structured filtering on top of vector similarity. Skip if you only need simple k-NN without metadata constraints—lighter tools exist. The Rust implementation and explicit query planner suggest serious engineering, not just another HNSW wrapper.