nmslib/hnsw
Hnsw
该项目是一个 C++ 向量相似度引擎，实现了分层可导航小世界（HNSW）算法。它提供了一个仅包含头文件的库，用于在高维空间中执行近似最近邻搜索，并附带 Python 绑定，将这些索引和搜索功能暴露给数据科学环境。
该引擎利用多层图结构平衡搜索速度和准确性，实现实时嵌入检索和高维相似度搜索。它支持自定义距离度量，以计算各种数学空间中向量之间的相似度。
该系统涵盖了广泛的索引操作，包括增量向量插入以及在不重建索引的情况下标记元素为已删除的能力。它还支持多向量搜索、搜索结果过滤以及基于磁盘的持久化，以保存和重新加载索引状态。
Features
- Vector Similarity Search - Provides a high-performance engine for similarity search on high-dimensional vector embeddings.
- Low-Latency Vector Retrieval - Enables low-latency retrieval of the most relevant vectors from large datasets in real-time.
- Approximate Nearest Neighbor Search - Finds the closest vectors in high-dimensional space by balancing lookup speed and accuracy.
- Proximity Graph Indexes - Uses a multi-layered proximity graph to efficiently navigate and retrieve approximate nearest neighbors.
- Vector Indexing - Organizes high-dimensional embeddings into optimized indexes to support efficient semantic retrieval.
- HNSW Indexes - Implements the Hierarchical Navigable Small World algorithm for fast approximate nearest neighbor search.
- Vector Distance Metrics - Provides pluggable distance functions to calculate similarity between high-dimensional vectors in various mathematical spaces.
- Disk-Persistent Data Structures - Implements the ability to serialize the in-memory graph structure to disk for session recovery.
- Incremental Indexing Engines - Allows for the insertion and modification of vectors without rebuilding the entire index from scratch.
- Python Vector Index Wrappers - Exposes high-performance C++ vector indexing and search capabilities to Python environments.
- Search Result Filtering - Restricts search results to specific subsets of elements using labels or boolean logic.
- Incremental Vector Sync - Supports adding new vectors to the index incrementally without requiring a full rebuild.
- Header-Only Libraries - Distributed as a header-only C++ library to simplify integration and enable compiler optimizations.
- Python-C++ Bindings - Provides Python bindings for high-performance C++ vector indexing and search operations.
- Radius Vector Searches - Supports executing queries using distance thresholds to find all neighbors within a specific radius.
- Similarity Search - Graph-based method for efficient approximate nearest neighbor search.
