---
格式版本: 2
标题: "The Commercial Tax: Rent-vs-Own Blind Spots in Multi-Hop Retrieval Benchmarks"
原文链接: "https://arxiv.org/abs/2608.16096"
发布日期: "2026-08-17"
发布时间校准状态: "found"
发布时间需复核: "否"
发布时间来源: "llm:scrape:strict_html_body"
发布时间证据: "[Submitted on 17 Aug 2026]"
发布时间校准原因: "arXiv 页面明确标注提交日期为 2026-08-17，该日期即为论文发布时间，优先于其他元数据或正文信息。"
发布时间校准置信度: "1"
发布时间候选数量: 26
发布时间严格候选数量: 9
发布时间原页读取状态: "source template page reused from URL open"
发布时间未找到原因: ""
发布时间校准时间: "2026-08-20T14:35:33+08:00"
发布时间仲裁状态: "confirmed"
发布时间仲裁尝试次数: 1
发布时间仲裁耗时毫秒: 17815
发现时间: "2026-08-20T14:32:53+08:00"
入库时间: "2026-08-20T06:35:51.105Z"
来源平台: "arXiv 学术论文搜索"
搜索渠道: "source_template"
搜索词: "https://arxiv.org/search/?query=NVIDIA&searchtype=all"
匹配关键词:
  - "performance"
相关厂家:
  - "NVIDIA"
  - "Microsoft"
相关专家:
  []
内容类型: "网页"
抓取工具: "Free Fetch + Defuddle"
清洗工具: "Defuddle Markdown + Defuddle/Readability 正文提取"
原始附件:
  []
AI优质: "否"
AI打分: 5
AI分档: "非优质"
AI质检状态: "不通过"
AI打分理由: "本文是信息检索领域的多跳检索基准论文，涉及嵌入模型许可与成本，仅提及NVIDIA品牌，与超节点/AI Rack/机柜级AI基础设施完全无关，属于明显的主题不相关。"
AI质检模型: "ali-deepseek-v4-flash"
AI质检时间: "2026-08-20T14:36:11+08:00"
AI主题相关性: 0
AI来源权威性: 5
AI新颖性: 0
AI技术细节: 0
AI商业部署信号: 0
AI完整性: 0
AI摘要: "该文审计多跳检索基准，指出其忽略商业许可和构建成本，且多套领先系统依赖非商业许可的NV-Embed-v2却未披露。"
AI摘要模型: "ali-deepseek-v4-flash"
AI摘要时间: "2026-09-07T03:24:55.734Z"
采集批次: "2026年8月20日14点19分32秒"
采集批次ID: "20260820-141932-079"
去重键: "https://arxiv.org/abs/2608.16096"
---

## Computer Science > Information Retrieval

## Title:The Commercial Tax: Rent-vs-Own Blind Spots in Multi-Hop Retrieval Benchmarks

Authors:[Luis M. Sanchez](https://arxiv.org/search/cs?searchtype=author&query=Sanchez,+L+M), [Kosrow Dehnad](https://arxiv.org/search/cs?searchtype=author&query=Dehnad,+K)

[View PDF](https://arxiv.org/pdf/2608.16096) [HTML (experimental)](https://arxiv.org/html/2608.16096v1)

> Abstract:Enterprises connect language models to their own data through retrieval. The benchmarks that rank multi-hop retrieval systems leave out two facts a buyer needs before a published number can be used: whether the retrieval backbone may be deployed commercially, and what it costs to build. On licensing: the field's dense-retrieval anchor, NV-Embed-v2, is licensed cc-by-nc-4.0. Of the four leading MuSiQue systems we audit (HippoRAG-2, PropRAG, SAG, KET-RAG), three depend on it for their best numbers and none says so. On performance: we measure thirteen embedders from eight makers on one identical MuSiQue harness with bootstrap confidence intervals throughout. Until mid-2026 there was a real commercial tax: the best commercially-licensed embedder trailed the anchor by 2.31 Recall@5 points (95% CI \[0.91, 3.71\], p=0.001). NVIDIA's Nemotron-3-Embed-8B, released 2026-07-16, has closed it: +0.24 at Recall@5 (95% CI \[-0.94, +1.43\], p=0.69), -0.58 at Recall@10 (p=0.28). It matches the anchor, does not beat it, and is the only entrant that is commercially licensed, free to self-host, and indistinguishable from the anchor; every other entrant meeting the first two conditions sits 5.2 to 14.6 points below. The durable finding is the paid-versus-free divide: API embedders charge per token on every re-index, self-hosted ones charge nothing. On cost: three of five audited systems (adding Microsoft's GraphRAG) do not disclose indexing cost, and the only published GraphRAG dollar figures span 11x inside one third-party paper (USD 2.30 vs USD 24.94 to index a 5.64 MB corpus once); extrapolated to 1 TB that undisclosed choice separates roughly USD 428K from $4.6M. Our cost model keeps one-time embedding apart from recurring answering: at 1 TB, embedding sits 7.5x-900x below graph construction, and a year of answering at 10,000 queries/day sits 350x or more below it.

| Comments: |  |
| --- | --- |
| Subjects: | Information Retrieval (cs.IR); Computation and Language (cs.CL) |
| Cite as: | [arXiv:2608.16096](https://arxiv.org/abs/2608.16096) \[cs.IR\] |
|  | (or [arXiv:2608.16096v1](https://arxiv.org/abs/2608.16096v1) \[cs.IR\] for this version) |
|  | [https://doi.org/10.48550/arXiv.2608.16096](https://doi.org/10.48550/arXiv.2608.16096) |

## Submission history

From: Luis M Sanchez \[[view email](https://arxiv.org/show-email/710f142a/2608.16096)\]  
**\[v1\]** Mon, 17 Aug 2026 04:37:01 UTC (142 KB)

[Which authors of this paper are endorsers?](https://arxiv.org/auth/show-endorsers/2608.16096) | Disable MathJax ([What is MathJax?](https://info.arxiv.org/help/mathjax.html))