Show HN: Agent Composer – 用于火箭科学（以及其他难题）的 AI 智能体
1 分•作者: jayc481•17 天前
大家好，我是 Contextual AI 的 Jay（网址：https://contextual.ai/）。
我们一直在构建一个面向技术行业的 AI 智能体平台，主要针对半导体、航空航天、制造业等领域。Agent Composer 是我们新推出的强大可视化构建器和运行时，用于创建能够基于技术文档、日志和规范进行推理的智能体。
我们解决的问题是：通用 AI 在复杂的的技术任务上表现不佳。这并非因为模型本身能力不足，而是因为它们无法访问正确的上下文信息（数据表、测试日志、流程规范、机构知识）。
Agent Composer 的功能：
* 三种创建智能体的方式：预构建模板、自然语言描述或空白画布
* 可视化拖放式构建器，提供无代码体验，同时为开发人员提供 YAML 配置
* 混合工作流程：将确定性步骤（合规性检查、验证）与动态推理（根本原因分析、研究）相结合
* 基于您的数据，并提供完整归因
我们在此过程中学到的：
* 解析比人们想象的更重要。包含表格、图表和交叉引用的技术 PDF 会让大多数现成的解析器崩溃。我们构建了自己的解析器。
* 检索精度至关重要。基本的向量搜索可以解决 70% 的问题；剩下的 30% 需要混合检索、重新排序和查询重构。这最后的 30% 区分了“炫酷演示”和“真正有用”。
* 企业需要控制。纯粹的自主智能体会让合规团队感到担忧。在同一个工作流程中混合确定性和动态步骤的能力，是对客户反馈的直接回应。
以下是一些供您探索的链接：
* 产品快速入门指南：https://docs.contextual.ai/quickstarts/agent-composer
* 我们构建的有趣的火箭科学演示：https://demo.contextual.ai/
* 博客：https://contextual.ai/blog/introducing-agent-composer
* 免费帐户注册链接：https://app.contextual.ai/?signup=1
很乐意深入探讨架构、检索策略或经验教训。您有什么问题或反馈吗？
查看原文
Hi HN, Jay from Contextual AI (<a href="https://contextual.ai/" rel="nofollow">https://contextual.ai/</a>) here.<p>We've been building a platform for AI agents focused on technical industries—semiconductors, aerospace, manufacturing, etc. Agent Composer is our powerful new visual builder and runtime for creating agents that reason over technical documentation, logs, and specs.<p>The problem we solved: General-purpose AI fails on complex technical tasks. Not because the models aren't capable, but because they don't have access to the right context (datasheets, test logs, process specs, institutional knowledge).<p>What Agent Composer does:<p>- Three ways to create agents: pre-built templates, natural language description, or blank canvas<p>- Visual drag-and-drop builder for a no-code experience and YAML configs available for developers<p>- Hybrid workflows: combine deterministic steps (compliance checks, validation) with dynamic reasoning (root cause analysis, research)<p>- Grounded in your data with full attribution<p>What we learned building this:<p>Parsing matters more than people think. Technical PDFs with tables, figures, and cross-references break most off-the-shelf parsers. We built our own.<p>Retrieval precision is everything. Basic vector search gets you 70% of the way; the last 30% requires hybrid retrieval, reranking, and query reformulation. That last 30% is the difference between "neat demo" and "actually useful."<p>Enterprises need control. Pure autonomous agents scare compliance teams. The ability to mix deterministic and dynamic steps in one workflow was a direct response to customer feedback.<p>Here are some links for you to explore:<p>- Product quick-start guide: <a href="https://docs.contextual.ai/quickstarts/agent-composer" rel="nofollow">https://docs.contextual.ai/quickstarts/agent-composer</a><p>- Fun rocket science demo we built: <a href="https://demo.contextual.ai/" rel="nofollow">https://demo.contextual.ai/</a><p>- Blog: <a href="https://contextual.ai/blog/introducing-agent-composer" rel="nofollow">https://contextual.ai/blog/introducing-agent-composer</a><p>- Free account signup link: <a href="https://app.contextual.ai/?signup=1" rel="nofollow">https://app.contextual.ai/?signup=1</a><p>Happy to go deep on architecture, retrieval strategies, or lessons learned. What questions or feedback do you have?