Qdrant

Performs approximate nearest neighbor (ANN) search on dense vectors using Hierarchical Navigable Small World (HNSW) graphs, enabling sub-millisecond retrieval at scale. Vectors are indexed in-memory with configurable M and ef parameters controlling graph connectivity and search quality tradeoffs. Supports batch queries and single-vector lookups with configurable result limits and score thresholds.

Executes unified search across both dense embeddings (semantic) and sparse vectors (keyword/BM25), fusing results using configurable weighting strategies. Sparse vectors are generated via SPLADE++, miniCOIL, or BM25 algorithms and indexed separately from dense vectors. Results from both indices are merged using RRF (Reciprocal Rank Fusion) or weighted linear combination, enabling queries to match both semantic meaning and exact keywords.

Defines collection schema specifying vector dimensionality, distance metric (cosine, dot product, Euclidean), payload field types, and indexing strategy. Schema is enforced on insert; vectors not matching schema are rejected. Supports schema evolution (adding new fields) without reindexing. Distance metrics are configurable per collection, enabling different similarity measures for different use cases.

Supports batch insert, update, and delete operations on multiple vectors in a single request, with all-or-nothing transactional semantics. Batch operations are more efficient than individual requests (10-100x throughput improvement). Supports upsert (insert-or-update) for idempotent operations. Batch size limits are configurable.

Exposes vector search functionality via both REST API (HTTP/JSON) and gRPC (binary protocol). REST API is suitable for web applications and simple integrations; gRPC is optimized for high-throughput and low-latency scenarios. Language-specific SDKs are available for Python, JavaScript/TypeScript, Rust, Go, and Java, providing idiomatic interfaces and automatic serialization. SDKs handle connection pooling, retries, and error handling.

Fully managed Qdrant deployment on AWS, GCP, or Azure with automatic vertical and horizontal scaling based on resource utilization. Includes automated backups, monitoring, alerting, and 99.5% (standard) or 99.9% (premium) uptime SLA. Eliminates operational overhead of self-hosted deployments. Pricing is usage-based (compute and storage).

Qdrant can be deployed as a Docker container or on Kubernetes clusters, enabling self-hosted deployments on any infrastructure (on-premises, private cloud, hybrid cloud). Includes Helm charts for Kubernetes deployment and Docker Compose examples for single-node setups. Supports persistent storage via volumes and external object storage for snapshots. No licensing fees for self-hosted deployments.

Applies complex metadata filters during vector search using a JSON-based query language supporting nested objects, arrays, text matching, numeric ranges, geospatial bounding boxes, and has_vector predicates. Filters are evaluated during HNSW traversal (one-stage filtering), not post-retrieval, reducing memory overhead. Supports AND/OR/NOT boolean logic and arbitrary nesting depth.

Stores and indexes multiple dense vectors per data point (e.g., multiple embeddings for different modalities or text chunks), enabling retrieval of the same logical entity from multiple vector spaces. Each vector is indexed independently in HNSW, but queries return the parent point with all associated vectors. Supports named vector fields for explicit multi-modal or multi-representation scenarios.

Reduces vector memory footprint and accelerates search by quantizing dense vectors to lower precision (int8 scalar quantization, int4 product quantization, binary quantization). Quantized vectors are indexed in HNSW with minimal recall loss (typically 1-5% depending on quantization type). Supports mixed-precision indexing where some vectors are quantized and others remain full-precision.

Newly inserted vectors are immediately searchable without explicit index rebuild or refresh cycles. Vectors are added to HNSW graph in real-time, and search queries see the latest data within milliseconds. No batch indexing delays or eventual consistency windows. Supports both single-vector and batch insert operations with transactional semantics.

Re-scores initial HNSW search results using advanced ranking strategies including ColBERT late interaction models, Maximum Marginal Relevance (MMR) for diversity, and custom scoring functions. Reranking is applied post-retrieval on the top-K candidates, enabling expensive ranking models without full-collection scan. Supports combining multiple ranking signals (semantic similarity, diversity, business metrics).

Distributes vector collections across multiple nodes using consistent hashing and shard-based partitioning. Each shard maintains its own HNSW index and is replicated for fault tolerance. Queries are routed to relevant shards, and results are merged. Supports dynamic shard rebalancing and automatic failover. Enables scaling to billions of vectors across commodity hardware.

Creates consistent snapshots of collection state at specific points in time, enabling recovery to any snapshot without data loss. Snapshots are stored locally or in cloud object storage (S3, GCS, Azure Blob). Supports incremental snapshots to reduce storage overhead. Enables disaster recovery, data migration, and A/B testing of different vector versions.

Manages API authentication using vector-scoped API keys that can be restricted to specific collections, read-only vs read-write operations, and IP address ranges. Keys are generated and revoked via API or dashboard. Supports role-based access control (RBAC) for multi-tenant deployments. Enterprise tier includes audit logging of all API key usage.