qdrant

Implements Hierarchical Navigable Small World (HNSW) graph indexing for sub-linear time complexity nearest neighbor queries across dense vector spaces. The implementation uses a multi-layer graph structure where each layer is a navigable small world graph, enabling efficient approximate search by starting from the top layer and progressively descending. Supports configurable M (max connections per node) and ef (search expansion factor) parameters to tune the recall-latency tradeoff, allowing users to balance query speed against result accuracy without re-indexing.

Enables simultaneous search across dense vectors (via HNSW) and sparse vectors (via inverted indices) with configurable weighted combination of results. The system maintains separate index structures for dense and sparse vectors within each segment, executes parallel searches, and merges results using a weighted scoring function that combines dense similarity scores with sparse BM25-style relevance scores. This allows semantic search (dense) and keyword matching (sparse) to be unified in a single query without requiring separate round-trips.

Implements write-ahead logging (WAL) to ensure data durability and consistency, with configurable fsync policies to balance durability against write latency. Each write operation is logged to disk before being applied to in-memory indices, enabling recovery from crashes without data loss. Fsync policies range from immediate (fsync after every write, highest durability but highest latency) to batched (fsync every N writes, lower latency but higher data loss risk). WAL is used for both point-in-time recovery and segment compaction consistency.

Supports batch operations (upsert, delete, update) that are applied atomically within a single request, ensuring all operations in the batch succeed or all fail together. Batch operations are processed through the update pipeline and applied to segments in a single transaction, maintaining consistency across multiple point updates. This enables efficient bulk loading and updates without requiring separate requests for each operation.

Provides a lightweight embedded library (Qdrant Edge) that runs vector search directly on edge devices (mobile, IoT, embedded systems) without requiring a server connection. The library is a minimal Rust implementation of Qdrant's core search functionality (HNSW search, filtering, quantization) compiled to WebAssembly or native binaries for edge platforms. Edge library supports pre-built indices that are downloaded from the server and cached locally, enabling offline search with periodic synchronization.

Provides optional inference service integration that generates embeddings from raw text/images using configurable embedding models (e.g., OpenAI, Hugging Face, local models). The inference service is decoupled from the vector database; clients can use it to generate embeddings before inserting into Qdrant, or Qdrant can be configured to call the inference service during upsert operations. This enables end-to-end workflows where raw documents are inserted and embeddings are generated automatically.

Provides structured filtering on document metadata (payloads) using field-specific index types (keyword, integer range, geo-spatial, full-text) that are selected automatically or manually based on field type and query patterns. Each field maintains its own index structure (e.g., B-tree for ranges, inverted index for keywords, R-tree for geo) stored alongside vector indices in segments. Filters are applied during search to prune candidates before distance computation, reducing the search space and improving query latency for selective filters.

Reduces memory footprint and improves search speed by quantizing dense vectors to lower precision (int8, uint8, or binary) while maintaining configurable recall through quantization-aware distance calculations. Supports both product quantization (PQ) and scalar quantization (SQ) approaches, where vectors are decomposed into subspaces or scaled to lower bit-widths. Quantized vectors are stored in segments alongside original vectors (or as the only copy), and distance computations use quantization-aware metrics that account for precision loss.

Distributes vector collections across multiple shards (horizontal partitioning) and maintains replica sets for fault tolerance, with automatic failover when shard replicas become unavailable. The system uses Raft consensus to maintain consistency across replicas and automatically detects peer failures through heartbeat monitoring. Queries are routed to available shard replicas, and if a primary replica fails, the system promotes a secondary replica without manual intervention. Shard transfers and resharding are orchestrated through the Raft-based consensus layer.

Organizes data within each shard into immutable segments that are created during writes and automatically compacted/optimized based on size and update patterns. Each segment contains vectors, indices (HNSW, field indices), and metadata stored in a columnar format optimized for sequential access. The segment lifecycle manager monitors segment sizes and fragmentation, triggering compaction when segments become too small or fragmented, merging multiple segments into larger optimized segments. This design enables efficient incremental updates without full index rebuilds while maintaining query performance.

Creates consistent snapshots of collections at specific points in time, capturing all segments, indices, and metadata needed to restore the collection to that exact state. Snapshots are stored as compressed archives containing segment data and can be transferred between nodes for recovery or cloning. The snapshot mechanism uses write-ahead logging to ensure consistency; snapshots capture the state after all writes up to a specific log position, enabling point-in-time recovery without data loss.

Exposes vector database operations through both REST (HTTP/JSON) and gRPC (Protocol Buffers) APIs with identical functionality, allowing clients to choose based on performance and integration requirements. REST API is built on actix-web framework and gRPC on tonic framework, both routing to the same underlying dispatcher and collection management layer. This dual-protocol approach enables easy integration with web applications (REST) while supporting high-performance services (gRPC) without maintaining separate code paths.

Offloads computationally intensive vector operations (distance calculations, HNSW graph traversal) to GPU when available, using CUDA for NVIDIA GPUs. GPU acceleration is transparent to clients; the system automatically detects GPU availability and routes eligible operations to GPU kernels while falling back to CPU for unsupported operations or when GPU is unavailable. Distance calculations benefit most from GPU acceleration (10-50x speedup for large batches), while HNSW traversal benefits less due to irregular memory access patterns.

Allows multiple collection names to point to the same underlying data, enabling zero-downtime updates by creating a new collection, indexing data, and atomically switching the alias to the new collection. Aliases are stored in the distributed consensus layer (Raft) and switches are atomic across the cluster. This pattern enables blue-green deployments where the old collection remains available until the new one is fully indexed, then traffic is switched via alias update.