Show HN: 通过人工反馈循环进行文档提取
1 分•作者: sails•8 个月前
关于该功能：
我们构建了一个工具，可以从复杂的管理财务文件和银行对账单中自动评估企业的健康状况。
为了使其可靠地运行，我们不得不对字段级别的结构化输出提示进行大量迭代。
这促使我们构建了一个功能，允许用户改进这些提示，并通过一些上下文工程，我们得到了我们称之为“反馈循环”的东西。
这是一个 2 分钟的演示 <a href="https://www.youtube.com/watch?v=ZDNlEZydoXU" rel="nofollow">https://www.youtube.com/watch?v=ZDNlEZydoXU</a>
工作原理（视频演示了以下步骤）：
1. 为您的提取创建目标表单（AI 帮助创建）
2. 上传文档（从不同的模型中选择）
3. 审查提取的质量（并检查 PDF 引用）
4. 如果有任何错误，请更正它们并在字段级别提供反馈
5. 一旦您觉得您已经看到了足够的错误并提供了足够的更正，请使用反馈工作流程来完善您的字段描述。
您可以在这里试用该应用程序 <a href="https://app.sea.dev/" rel="nofollow">https://app.sea.dev/</a>
（在提示优化之间至少进行 2-3 次合理的提取时，效果最佳）
反馈功能目前处于 Beta 测试阶段。如果您进行了一些更正，然后转到 <a href="https://app.sea.dev/feedback" rel="nofollow">https://app.sea.dev/feedback</a>，您将看到它们已准备好进行审查和完善。
我很乐意收到关于我们如何改进这个想法的反馈。该视频是一个玩具示例，但它已经上线并与一些早期的信用分析师客户一起使用，他们似乎很喜欢它。
查看原文
About the feature:<p>We built a tool for automatically assessing business health from complex management financial documents and bank statements.<p>In order to get this working reliably, we had to iterate extensively on the prompts for the structured output at the field level.<p>This led us to building a feature to allow users to improve them too, and with some context engineering, we have what we are calling “feedback loops”.<p>Here is a 2 min demo <a href="https://www.youtube.com/watch?v=ZDNlEZydoXU" rel="nofollow">https://www.youtube.com/watch?v=ZDNlEZydoXU</a><p>How it works (video runs through these steps):<p>1. Create a target form for your extraction (AI helps create it)<p>2. Upload a document (choose from different models)<p>3. Review the quality of the extraction (and check the PDF citations)<p>4. If there are any mistakes, correct them and give feedback at the field level<p>5. Once you feel like you've seen enough errors and provided enough corrections, use the Feedback workflow to refine your field descriptions.<p>You can try the app here <a href="https://app.sea.dev/" rel="nofollow">https://app.sea.dev/</a><p>(it works best when you do at least 2-3 reasonable extractions between prompt refinements)<p>The feedback feature is currently in beta. If you make some corrections and then go to <a href="https://app.sea.dev/feedback" rel="nofollow">https://app.sea.dev/feedback</a> you'll see them ready for review and refinement.<p>I would love to get feedback on how we might improve this idea. The video was a toy example but it is live and working with a few early credit analysts customers and they seem to like it.