ScaNN

ScaNN (Scalable Nearest Neighbors) is an open-source library developed by Google Research for efficient vector similarity search, particularly designed for large-scale nearest neighbor search applications in AI and machine learning. It is optimized for high-dimensional vector data, such as embeddings generated from text, images, or other modalities.

Project Source / More Info

Features

• Efficient Approximate Nearest Neighbor Search: Enables fast similarity search in large datasets of high-dimensional vectors, suitable for millions or billions of items.
• Anisotropic Vector Quantization: Implements a novel quantization technique that penalizes error in the direction parallel to the original vector, improving accuracy for maximum inner-product search (MIPS).
• Optimized for MIPS: Specifically designed to accelerate maximum inner-product search, a common operation in embedding-based retrieval tasks.
• High Performance: Outperforms other vector similarity search libraries on standard benchmarks (e.g., ann-benchmarks.com), achieving up to twice the query throughput for a given accuracy.
• Open Source: Available for direct installation via Pip, with source code and documentation on GitHub.
• Flexible Interfaces: Supports both TensorFlow and Numpy inputs for easy integration into various machine learning workflows.
• Scalable: Handles very large datasets, making it suitable for production-scale AI systems.

Category

SDKs & Libraries

Tags

open-source, ann, vector-search, ai

Pricing

ScaNN is open-source software and is available for free.

Links

• Official Blog Announcement
• GitHub Repository