Qdrant
Sur cette page
Qdrant Rust vector DB performance focus Apache 2.0.
Définition
Qdrant features : (1) Collections : analog to tables, defined with vector size + distance metric (Cosine, Dot, Euclidean, Manhattan). (2) Index : HNSW (Hierarchical Navigable Small World) optimized Rust implementation, integration AlmostHNSW recent improvements. (3) Quantization : Scalar Quantization (32-bit float -> 8-bit int4 4x compression), Binary Quantization (1-bit per dimension 32x compression, faster but accuracy loss), Product Quantization PQ. (4) Filtering : Payload-based filtering integrated index (Filterable Index), pre-filtering during HNSW traversal (vs post-filtering common alternatives). (5) Multi-tenancy : tenant isolation via collection separation ou payload-based. (6) Distributed : multi-node cluster support sharding + replication, Raft consensus, Kubernetes Helm chart. (7) Sparse Vectors : SPLADE, native support hybrid dense + sparse. (8) Performance : Rust efficiency, ~10x memory reduction vs Python Milvus equivalent, sub-10ms latency typical. Pricing : OSS free + Qdrant Cloud Free Tier (1GB storage) + Standard tier $0.05/GB-month + Hybrid Cloud BYOC. Customers : Twitch Search, Vivino, Bayer, HubSpot. $28M Series A 2024.
Les filtres de payload couvrent notamment les critères géographiques, les dates et les objets imbriqués. Qdrant gère plusieurs vecteurs par point pour le RAG multimodal et propose des SDK Python, JavaScript, Go et Rust.
Origine
Qdrant fondee 2021 a Berlin par Andrey Vasnetsov + Andre Zayarni ; Seed $7.5M 2022 ; Series A $28M janvier 2024 (Spark Capital lead) ; ~50000+ self-hosted instances ; ~3000 Qdrant Cloud customers 2024.
Exemple en contexte
Twitch Search utilise Qdrant ~100M+ video clips embedded via custom multi-modal model (vision + audio + text), semantic search clips ~50ms latency, Qdrant cluster ~12 Kubernetes nodes c5.4xlarge, Binary Quantization reduces memory 32x compared float32 baseline.