cuVS Documentation

cuVS provides a suite of preprocessing APIs designed to transform, compress, or project vectors before they are used in downstream search or analysis tasks. These tools help optimize storage requirements and computational efficiency.

Available preprocessing techniques include:

• Binary Quantizer: Compresses vectors into binary representations for compact storage and high-speed comparisons.
• PCA (Principal Component Analysis): Reduces dimensionality using linear projection while attempting to preserve maximum variance.
• Product Quantization (PQ): Splits vectors into subvectors and encodes each part using compact codebooks.
• Scalar Quantizer: Compresses each vector dimension independently using scalar quantization.
• Spectral Embedding: Generates lower-dimensional embeddings derived from graph structures.

cuVS Bench is a reproducible benchmarking tool designed for Approximate Nearest Neighbor (ANN) search implementations. It allows developers to compare performance across different hardware configurations, including GPU-to-GPU and GPU-to-CPU comparisons.

Key capabilities include:

• Comparing build time, search throughput, latency, and recall.
• Finding optimal parameter settings for recall buckets.
• Generating consistent plots for performance analysis.
• Identifying optimization opportunities across index parameters, build time, and search performance.
• Capturing index configurations that are reproducible across on-prem and cloud hardware.

NVIDIA cuVS is a GPU-accelerated library designed for vector search, including approximate and exact nearest neighbors, vector compression, and clustering. It provides both high-level end-to-end algorithms and low-level building blocks for constructing new algorithms.

Key capabilities include:

• Vector Search: Finding semantically similar items in embedding spaces (used in RAG, Recommender systems, and Computer Vision).
• Clustering & Visualization: Accelerating algorithms like K-means, UMAP, t-SNE, and HDBSCAN.
• Graph Construction: Converting dense vectors into nearest neighbor graphs for use with libraries like cuGraph or GraphBLAS.

The library is built on top of the RAPIDS RAFT library of high-performance machine learning primitives.

The brute_force Go package provides tools for building and searching a Brute Force K-Nearest Neighbors (KNN) index. This is useful for exact nearest neighbor searches on a dataset using a provided distance metric.

Key components include:

• BruteForceIndex: The core index type.
• CreateIndex(): To initialize an empty index.
• BuildIndex(): To populate an index with a dataset.
• SearchIndex(): To perform queries against the index.
• Close(): To release resources associated with the index.

cuVS can be used to inspect and maintain the quality of embedding models and datasets:

• Embedding Quality Inspection: Use nearest-neighbor examples to verify if a model correctly groups related items or if it encodes unwanted shortcuts. Batch searches can identify mislabeled examples, outliers, or hard negatives.
• Dataset Drift Tracking: Use clustering and graph analysis to monitor if new embeddings form new clusters or change neighbor relationships, which signals changes in data distribution or model behavior.

The cuVS Java API provides core abstractions for managing data on different memory locations and interacting with GPU resources. Key components include:

• Matrix Abstractions: Use CuVSMatrix as a general interface, or specific implementations like CuVSDeviceMatrix for GPU-resident data and CuVSHostMatrix for CPU-resident data.
• Resource Management: Manage GPU and library resources using CuVSResources and CuVSResourcesInfo.
• GPU Information: Access hardware details via GPUInfo and GPUInfoProvider.
• Service Providers: Interact with the cuVS ecosystem through CuVSProvider and CuVSServiceProvider.
• Error Handling: Catch LibraryException for errors occurring within the cuVS library during execution.