Q2D-Web
收藏资源简介:
Q2D-Web(Query2Doc-Web)是由Perplexity AI构建的大规模智能体检索基准,包含约1.9亿文档的网页语料库和约7万条经智能体重写的搜索查询,覆盖十种语言,采样自九个月的无个人识别信息生产流量。数据集创建过程包括查询采样、语料库构建、多源相关性标注(结合智能体引用、生产排序及LLM判断)以及子语料库采样以支持高效评估。该基准旨在评估生产级RAG系统中第一阶段检索器的深度召回能力,弥补现有基准在规模、查询类型和标注深度上的不足,为智能体检索研究提供更真实、更全面的测试平台。
Q2D-Web (Query2Doc-Web) is a large-scale intelligent retrieval benchmark developed by Perplexity AI. It comprises a web corpus with approximately 190 million documents and around 70,000 search queries intelligently rewritten by AI agents, covering ten languages and sampled from nine months of production traffic free of personally identifiable information (PII). The dataset creation process includes query sampling, corpus construction, multi-source relevance annotation (integrating AI agent citations, production ranking results, and LLM judgments), as well as sub-corpus sampling to support efficient evaluation. This benchmark is designed to evaluate the deep recall capability of first-stage retrievers in production-grade Retrieval-Augmented Generation (RAG) systems, addressing the shortcomings of existing benchmarks in terms of scale, query types, and annotation depth, and providing a more realistic and comprehensive testbed for intelligent retrieval research.

- 1Q2D-Web: A Large-Scale Benchmark for Retrieval in Agentic RAG SystemsPerplexity AI · 2026年



