Qdrant

Qdrant stores vectors together with arbitrary JSON payloads and searches them by similarity. What it removes is the split most teams start with, an approximate nearest neighbour index on one side and a database holding the metadata on the other, because filtering happens inside the same query that ranks the results.

The engine is written in Rust and exposes a REST API with an OpenAPI specification alongside a gRPC interface for production traffic. Official clients cover Go, Rust, JavaScript, Python, .NET and Java. A single docker run command starts a server, though that default deployment has no authentication and should be secured before it faces a network.

Search behavior is configurable well beyond plain nearest neighbour lookup.

• Dense, sparse and multivector search: semantic similarity, full-text matching and late interaction models such as ColBERT.
• Payload filtering: keyword, full-text, numeric range and geo conditions combined with should, must and must not clauses.
• Hybrid search: several vectors in one query, merged through fusion strategies such as Reciprocal Rank Fusion.
• Quantization and on-disk storage: cuts RAM use sharply and lets you trade search speed against precision.
• Distributed deployment: sharding and replication, with collections updated or resized without downtime.

Beyond the server there is Qdrant Edge, a build that runs inside the application process for offline or resource-constrained devices and synchronizes with a server later. Multitenancy, write-ahead logging, GPU-accelerated indexing and a web UI for exploring collections round out an Apache 2.0 licensed core.