Vector Index Benchmark for Embeddings (VIBE)
收藏资源简介:
VIBE是一个开源项目,旨在为近似最近邻搜索算法提供一个基准测试平台。它包含了使用现代密集嵌入模型创建基准数据集的流程,例如检索增强生成(RAG)。此外,VIBE还包含了来自不同分布的查询和语料库的分布外(OOD)数据集,以模拟真实世界的工作负载。VIBE对最先进的向量索引进行了全面的评估,在12个分布内和6个分布外数据集上对21种实现进行了基准测试。
VIBE is an open-source project designed to provide a benchmarking platform for approximate nearest neighbor search algorithms. It includes the workflow for constructing benchmark datasets using modern dense embedding models, such as Retrieval-Augmented Generation (RAG). Furthermore, VIBE contains out-of-distribution (OOD) datasets composed of queries and corpora from varied distributions to simulate real-world workloads. VIBE conducts comprehensive evaluations of state-of-the-art vector indexes, benchmarking 21 implementations across 12 in-distribution and 6 out-of-distribution datasets.

- 1VIBE: Vector Index Benchmark for Embeddings赫尔辛基大学, 阿尔托大学, 帕多瓦大学, 哥本哈根信息技术大学 · 2025年



