Perplexity 发布 Q2D-Web 基准，用于评估代理式 RAG 系统中的检索性能

报道来源 [9]

1. X — Perplexity TIER_1 English(EN) · perplexity_ai · 2026-09-09 20:21

   如需请求评估，请通过我们的评估请求表单提交一个公开可用的 Hugging Face 检索模型：https://t.co/yGvVSxnrC3

   To request an evaluation, submit a publicly available Hugging Face retrieval model through our evaluation request form: https://t.co/yGvVSxnrC3 The full technical report with methodology and findings is available here: https://t.co/G0AKX2N8r5

2. X — Perplexity TIER_1 English(EN) · perplexity_ai · 2026-09-09 20:21

   我们使用Recall@1000作为主要指标，在三个相关性数据集上评估了13种检索模型。

   We evaluated 13 retrieval models across three relevance sets, using Recall@1000 as the primary metric. pplx-embed-v1-4b leads Web Ranking (65.73) and Combined (69.11), while Nemotron-3-Embed-8B leads Citation (61.68). https://t.co/z81c25TGMx

3. X — Perplexity TIER_1 English(EN) · perplexity_ai · 2026-09-09 20:21

   互惠排名融合（RRF）子采样仅使用31.7%的文档即可在Combined Recall@1000上保留全语料库模型排名。

   Reciprocal rank fusion (RRF) subsampling preserves the full-corpus model ranking on Combined Recall@1000 using only 31.7% of the documents. For pplx-embed-v1-4b, this reduces evaluation from 4,608 to roughly 1,500 H200 GPU-hours. https://t.co/ZPTvJ1xYoN

4. X — Perplexity TIER_1 English(EN) · perplexity_ai · 2026-09-09 20:21

   Q2D-Web 使用代理引用、生产网络排名以及 LLM 判断扩展的组合集。

   Q2D-Web uses agent citations, production web rankings, and a combined set expanded with LLM judgments. These three relevance sets reduce false negatives and reliance on a single labeling pipeline, while testing how relevance definitions affect model performance. https://t.co/g6m…

5. X — Perplexity TIER_1 English(EN) · perplexity_ai · 2026-09-09 20:21

   该语料库结合了每个查询的前 5,000 个生产检索结果，并使用 MinHash-LSH 进行去重。

   The corpus combines the top 5,000 production retrieval results per query, deduplicated with MinHash-LSH. Each document is a plausible match for at least one query, including difficult distractors that match the topic but miss a required date, entity, or version. https://t.co/6XK…

6. X — Perplexity TIER_1 English(EN) · perplexity_ai · 2026-09-09 20:21

   Q2D-Web 涵盖编程、法律、健康、科学和金融，以及消费品、旅游、娱乐和本地信息。查询跨越十种语言

   Q2D-Web covers programming, law, health, science, and finance alongside consumer goods, travel, entertainment, and local information. Queries span ten languages, with English accounting for 65.8%. https://t.co/GdUEviYaGu

7. X — Perplexity TIER_1 English(EN) · perplexity_ai · 2026-09-09 20:21

   Q2D-Web 借鉴了九个月来收集的、涵盖十种语言和数十个领域的 23,000 次不含 PII 的生产搜索。

   Q2D-Web draws on 23,000 PII-free production searches across ten languages and dozens of domains, collected over nine months. Agents reformulate user requests into primary and support queries, each evaluated independently with its own relevance judgments. https://t.co/xUI4OxiBcr

8. X — Perplexity TIER_1 English(EN) · perplexity_ai · 2026-09-09 20:21

   真实的检索评估需要大型语料库和查询集，并进行深度相关性判断以减少假阴性。

   Realistic retrieval evaluation requires large corpora and query sets, with deep relevance judgments to reduce false negatives. Q2D-Web combines 190M web documents, 69,721 agent-reformulated queries, and 99.6 positive judgments per query on average in its combined set. https://t.…

9. X — Perplexity TIER_1 English(EN) · perplexity_ai · 2026-09-09 20:21

   我们正在推出 Q2D-Web (Query2Doc-Web)，一个用于评估代理式 RAG 系统中检索的基准和公开排行榜。

   We're introducing Q2D-Web (Query2Doc-Web), a benchmark and public leaderboard for evaluating retrieval in agentic RAG systems. Q2D-Web tests how embedding models perform on large-scale web search using agent-reformulated search queries. Read more: https://t.co/s476SxkE1L