HackerNews中文版

大家好，最近我一直在研究人工智能代理，并注意到一个普遍的烦恼：大多数AI绘图工具生成的是静态图像（PNG/JPG）。如果AI在某个细节上出错，你就无法轻易修复——必须重新生成整个图像。我想要一个能够生成可编辑代码并进行交互渲染的工具，因此我构建了DeepDiagram。这是一个开源的网络平台，采用多代理架构。它不是通过单一的提示尝试完成所有任务，而是通过路由器分析你的意图，并将其指向一个专业代理（例如，Draw.io专家、数据可视化专家或流程图专家）。演示思维导图与数据图表（ECharts）： Draw.io生成（基于XML）： Mermaid图表：技术栈编排：使用LangGraph（Python）进行ReAct循环和状态管理。前端：React 19、Vite、TailwindCSS、Zustand。后端：FastAPI（Python 3.10+）。流媒体：使用SSE（服务器推送事件）进行实时预览（打字机效果）。渲染引擎：React Flow、Apache ECharts、Mermaid.js和Draw.io（Atlas主题）。主要特性 Draw.io代理：生成兼容的XML。你可以直接在画布上编辑结果。数据图表代理：粘贴原始数据（CSV/JSON/文本），它会使用ECharts进行可视化（柱状图/折线图/饼图）。分支与时间旅行：由于AI并不完美，我实现了一个类似Git的分支系统。你可以“重试”一条消息，创建一个新分支，或返回到之前的状态查看代理的思考过程。自托管：支持Docker Compose。与OpenAI或DeepSeek V3配合良好（强烈推荐，性价比高）。链接 GitHub（AGPL-3.0）：https://github.com/twwch/DeepDiagram 它仍处于早期阶段，但我非常希望听到你的反馈或看到你的PR！

查看原文

Hi everyone,I've been working on AI Agents recently and noticed a common frustration: most AI diagramming tools generate static images (PNG/JPG). If the AI gets one detail wrong, you can't easily fix it—you have to regenerate the whole thing.I wanted a tool that generates editable code and renders it interactively. So I built DeepDiagram.It’s an open-source web platform that uses a Multi-Agent architecture. Instead of a single prompt trying to do everything, a Router analyzes your intent and directs it to a specialized agent (e.g., a Draw.io expert, a Data Visualization expert, or a Flowchart expert).Demos Mind Map & Data Charts (ECharts):Draw.io Generation (XML based):Mermaid Diagrams:Tech Stack Orchestration: LangGraph (Python) for the ReAct loop and state management.Frontend: React 19, Vite, TailwindCSS, Zustand.Backend: FastAPI (Python 3.10+).Streaming: Uses SSE (Server-Sent Events) for real-time preview (typewriter effect).Rendering Engines: React Flow, Apache ECharts, Mermaid.js, and Draw.io (Atlas theme).Key Features Draw.io Agent: Generates compatible XML. You can edit the result directly in the canvas.Data Chart Agent: Paste raw data (CSV/JSON/Text), and it visualizes it using ECharts (Bar/Line/Pie).Branching & Time Travel: Since AI isn't perfect, I implemented a Git-like branching system. You can "retry" a message, creating a new branch, or navigate back to previous states to see the agent's thought process.Self-Hosted: Docker Compose support included. Works great with OpenAI or DeepSeek V3 (highly recommended for cost/performance).Links GitHub (AGPL-3.0): https://github.com/twwch/DeepDiagramIt's still in the early stages, but I'd love to hear your feedback or see your PRs!

开源AI图表生成器（Draw.io，Mermaid）