0/6 主要航空航天文档门户已准备好支持人工智能代理
2 分•作者: priyanshu-j•3 个月前
我花了一些时间构建 AeroScore，这是一个评估主要文档门户对 AI 代理支持程度的评分引擎。简而言之，在我第一批评估的 6 个门户中，没有一个准备充分。
该方法论和标准基于 AFDocs Spec（一个用于代理就绪文档的开源清单），并增加了一个针对 PDF 繁重度的特定扩展，因为航空航天领域非常重视 PDF。每个网站根据 llms.txt 的存在、robots.txt 以及 AFDocs 存储库和 aeroscore 网页上列出的许多其他标准进行 0-100 分的评分。
主要发现：
- 0/6 网站有 llms.txt
- 0/6 网站支持 URL 变体
- 最佳分数是 ICAO，得分为 69/100
主要的航空航天文档门户尚未为 AI 代理做好准备，这一点很重要，因为它意味着该领域尚未充分发挥效率。AI 代理能够以预期时间的一小部分完成跨深度文档门户的数据分析，这已经彻底改变了许多领域。航空航天领域缺乏这一点意味着生产力的损失。
排行榜已上线：https://aeroscore.vercel.app
AFDocs 存储库地址：https://github.com/agent-ecosystem/afdocs
我是一名即将入学的航空航天工程本科生，对 AI 如何在我的领域中被实施以及可以如何实施感兴趣。这是 aeroscore 的 v1 版本，还比较粗糙，我非常感谢您对该方法论的反馈。在未来的版本中，我将使用改进后的评分标准对更多网站进行评分。
查看原文
I spent some time building AeroScore, a scoring engine that evaluates how well major documentation portals support AI agents. In short, none of the 6 that I evaluated in this first batch are sufficiently prepared.<p>The methodology and criteria are based on the AFDocs Spec (an open-source checklist for agent-ready documentation) and one specific extension for PDF-heaviness since aerospace as a field places so much importance on PDFs. Each site gets scored 0-100 across criteria like llms.txt presence, robots.txt, and many others that can be found both on the AFDocs repo and at the aeroscore webpage.<p>Key Findings:<p>-0/6 have an llms.txt<p>-0/6 support url variants<p>-the best score was ICAO with a 69/100<p>The fact that major aerospace documentation portals are not AI agent ready yet is important because it means the field is not being fully efficient. AI agents allow the analysis of data across deep documentation portals to occur in fractions of the expected time, and this has revolutionized many fields. The lack of this in aerospace means a loss of productivity.<p>The leaderboard is live at: https://aeroscore.vercel.app<p>The AFDocs repo is at: https://github.com/agent-ecosystem/afdocs<p>I'm an incoming undergraduate aerospace engineering student interested in how AI is and can be implemented in my field. This is v1 of aeroscore, it's pretty rough and I'd appreciate feedback on the methodology greatly. In future versions, I'll be scoring more sites with a refurbished rubric.