https://xiongneng.me/tags/langgraph/飞污熊小站Hugo 0.168.0 & FixIt v0.4.6-20260512073637-464c4659zh-CNyidao620@163.com (XiongNeng)yidao620@163.com (XiongNeng)XiongNengSat, 17 Oct 2026 14:40:13 +0000https://xiongneng.me/posts/ai/langchain/langchain-langgraph-12-travel-assistant/Mon, 12 Oct 2026 23:30:00 +0800yidao620@163.com (XiongNeng)https://xiongneng.me/posts/ai/langchain/langchain-langgraph-12-travel-assistant/AI<p><img loading="lazy" src='https://static.xiongneng.me/langchain-langgraph-cover-20261017205458.png' alt="https://static.xiongneng.me/langchain-langgraph-cover-20261017205458.png"></p> <h2 class="heading-element" id="学习进度"><span>学习进度</span> <a href="#%e5%ad%a6%e4%b9%a0%e8%bf%9b%e5%ba%a6" class="heading-mark"> <svg class="octicon octicon-link" viewBox="0 0 16 16" version="1.1" width="16" height="16" aria-hidden="true"><path d="m7.775 3.275 1.25-1.25a3.5 3.5 0 1 1 4.95 4.95l-2.5 2.5a3.5 3.5 0 0 1-4.95 0 .751.751 0 0 1 .018-1.042.751.751 0 0 1 1.042-.018 1.998 1.998 0 0 0 2.83 0l2.5-2.5a2.002 2.002 0 0 0-2.83-2.83l-1.25 1.25a.751.751 0 0 1-1.042-.018.751.751 0 0 1-.018-1.042Zm-4.69 9.64a1.998 1.998 0 0 0 2.83 0l1.25-1.25a.751.751 0 0 1 1.042.018.751.751 0 0 1 .018 1.042l-1.25 1.25a3.5 3.5 0 1 1-4.95-4.95l2.5-2.5a3.5 3.5 0 0 1 4.95 0 .751.751 0 0 1-.018 1.042.751.751 0 0 1-1.042.018 1.998 1.998 0 0 0-2.83 0l-2.5 2.5a1.998 1.998 0 0 0 0 2.83Z"></path></svg> </a> </h2><p>最后一篇，把整套技术串成一个能出门用的东西，个人旅游助手。你说一句「十一去杭州玩两天，不吃辣，喜欢博物馆」，它查天气、翻攻略、结合你的偏好排出行程，存成 markdown 文件，顺手把你的偏好记进长期记忆，下次再规划时不用重新交代。</p>https://xiongneng.me/posts/ai/langchain/langchain-langgraph-11-knowledge-base/Sun, 11 Oct 2026 22:30:00 +0800yidao620@163.com (XiongNeng)https://xiongneng.me/posts/ai/langchain/langchain-langgraph-11-knowledge-base/AI<p><img loading="lazy" src='https://static.xiongneng.me/langchain-langgraph-cover-20261017203759.png' alt="https://static.xiongneng.me/langchain-langgraph-cover-20261017203759.png"></p> <h2 class="heading-element" id="学习进度"><span>学习进度</span> <a href="#%e5%ad%a6%e4%b9%a0%e8%bf%9b%e5%ba%a6" class="heading-mark"> <svg class="octicon octicon-link" viewBox="0 0 16 16" version="1.1" width="16" height="16" aria-hidden="true"><path d="m7.775 3.275 1.25-1.25a3.5 3.5 0 1 1 4.95 4.95l-2.5 2.5a3.5 3.5 0 0 1-4.95 0 .751.751 0 0 1 .018-1.042.751.751 0 0 1 1.042-.018 1.998 1.998 0 0 0 2.83 0l2.5-2.5a2.002 2.002 0 0 0-2.83-2.83l-1.25 1.25a.751.751 0 0 1-1.042-.018.751.751 0 0 1-.018-1.042Zm-4.69 9.64a1.998 1.998 0 0 0 2.83 0l1.25-1.25a.751.751 0 0 1 1.042.018.751.751 0 0 1 .018 1.042l-1.25 1.25a3.5 3.5 0 1 1-4.95-4.95l2.5-2.5a3.5 3.5 0 0 1 4.95 0 .751.751 0 0 1-.018 1.042.751.751 0 0 1-1.042.018 1.998 1.998 0 0 0-2.83 0l-2.5 2.5a1.998 1.998 0 0 0 0 2.83Z"></path></svg> </a> </h2><p>这一篇做一个完整的项目，个人知识库。把你散落的 markdown 笔记吃进本地向量库，之后随时向它提问，它自己决定检索几次、检索什么，回答里的每个关键结论都带 <code>[文件名#块号]</code> 引用标记，顺着标记能翻回库里的原文块。</p>https://xiongneng.me/posts/ai/langchain/langchain-langgraph-10-multi-agent/Sat, 10 Oct 2026 21:30:00 +0800yidao620@163.com (XiongNeng)https://xiongneng.me/posts/ai/langchain/langchain-langgraph-10-multi-agent/AI<p><img loading="lazy" src='https://static.xiongneng.me/langchain-langgraph-cover-20261017193538.png' alt="https://static.xiongneng.me/langchain-langgraph-cover-20261017193538.png"></p> <h2 class="heading-element" id="学习进度"><span>学习进度</span> <a href="#%e5%ad%a6%e4%b9%a0%e8%bf%9b%e5%ba%a6" class="heading-mark"> <svg class="octicon octicon-link" viewBox="0 0 16 16" version="1.1" width="16" height="16" aria-hidden="true"><path d="m7.775 3.275 1.25-1.25a3.5 3.5 0 1 1 4.95 4.95l-2.5 2.5a3.5 3.5 0 0 1-4.95 0 .751.751 0 0 1 .018-1.042.751.751 0 0 1 1.042-.018 1.998 1.998 0 0 0 2.83 0l2.5-2.5a2.002 2.002 0 0 0-2.83-2.83l-1.25 1.25a.751.751 0 0 1-1.042-.018.751.751 0 0 1-.018-1.042Zm-4.69 9.64a1.998 1.998 0 0 0 2.83 0l1.25-1.25a.751.751 0 0 1 1.042.018.751.751 0 0 1 .018 1.042l-1.25 1.25a3.5 3.5 0 1 1-4.95-4.95l2.5-2.5a3.5 3.5 0 0 1 4.95 0 .751.751 0 0 1-.018 1.042.751.751 0 0 1-1.042.018 1.998 1.998 0 0 0-2.83 0l-2.5 2.5a1.998 1.998 0 0 0 0 2.83Z"></path></svg> </a> </h2><p>单个 agent 解决一个完整任务，任务复杂到需要查资料、写稿、审稿三种角色各干一摊时，把它们排进同一张图，就是多智能体编排。这一篇讲的就是怎么排。</p>https://xiongneng.me/posts/ai/langchain/langchain-langgraph-09-persistence/Fri, 09 Oct 2026 19:40:00 +0800yidao620@163.com (XiongNeng)https://xiongneng.me/posts/ai/langchain/langchain-langgraph-09-persistence/AI<p><img loading="lazy" src='https://static.xiongneng.me/langchain-langgraph-cover-20261017190854.png' alt="https://static.xiongneng.me/langchain-langgraph-cover-20261017190854.png"></p> <h2 class="heading-element" id="学习进度"><span>学习进度</span> <a href="#%e5%ad%a6%e4%b9%a0%e8%bf%9b%e5%ba%a6" class="heading-mark"> <svg class="octicon octicon-link" viewBox="0 0 16 16" version="1.1" width="16" height="16" aria-hidden="true"><path d="m7.775 3.275 1.25-1.25a3.5 3.5 0 1 1 4.95 4.95l-2.5 2.5a3.5 3.5 0 0 1-4.95 0 .751.751 0 0 1 .018-1.042.751.751 0 0 1 1.042-.018 1.998 1.998 0 0 0 2.83 0l2.5-2.5a2.002 2.002 0 0 0-2.83-2.83l-1.25 1.25a.751.751 0 0 1-1.042-.018.751.751 0 0 1-.018-1.042Zm-4.69 9.64a1.998 1.998 0 0 0 2.83 0l1.25-1.25a.751.751 0 0 1 1.042.018.751.751 0 0 1 .018 1.042l-1.25 1.25a3.5 3.5 0 1 1-4.95-4.95l2.5-2.5a3.5 3.5 0 0 1 4.95 0 .751.751 0 0 1-.018 1.042.751.751 0 0 1-1.042.018 1.998 1.998 0 0 0-2.83 0l-2.5 2.5a1.998 1.998 0 0 0 0 2.83Z"></path></svg> </a> </h2><p>这一篇给图装上两个进阶器官。持久化执行，图跑一半进程死了，换个进程接着跑。人工协同，图跑到关键步骤停下来，把问题递给人，等一个答复再继续。这两件事共享同一个地基，checkpointer 在每个 super-step 落一个状态快照。</p>https://xiongneng.me/posts/ai/langchain/langchain-langgraph-08-langgraph/Thu, 08 Oct 2026 19:10:00 +0800yidao620@163.com (XiongNeng)https://xiongneng.me/posts/ai/langchain/langchain-langgraph-08-langgraph/AI<p><img loading="lazy" src='https://static.xiongneng.me/langchain-langgraph-cover-20261017184837.png' alt="https://static.xiongneng.me/langchain-langgraph-cover-20261017184837.png"></p> <h2 class="heading-element" id="学习进度"><span>学习进度</span> <a href="#%e5%ad%a6%e4%b9%a0%e8%bf%9b%e5%ba%a6" class="heading-mark"> <svg class="octicon octicon-link" viewBox="0 0 16 16" version="1.1" width="16" height="16" aria-hidden="true"><path d="m7.775 3.275 1.25-1.25a3.5 3.5 0 1 1 4.95 4.95l-2.5 2.5a3.5 3.5 0 0 1-4.95 0 .751.751 0 0 1 .018-1.042.751.751 0 0 1 1.042-.018 1.998 1.998 0 0 0 2.83 0l2.5-2.5a2.002 2.002 0 0 0-2.83-2.83l-1.25 1.25a.751.751 0 0 1-1.042-.018.751.751 0 0 1-.018-1.042Zm-4.69 9.64a1.998 1.998 0 0 0 2.83 0l1.25-1.25a.751.751 0 0 1 1.042.018.751.751 0 0 1 .018 1.042l-1.25 1.25a3.5 3.5 0 1 1-4.95-4.95l2.5-2.5a3.5 3.5 0 0 1 4.95 0 .751.751 0 0 1-.018 1.042.751.751 0 0 1-1.042.018 1.998 1.998 0 0 0-2.83 0l-2.5 2.5a1.998 1.998 0 0 0 0 2.83Z"></path></svg> </a> </h2><p>LangGraph 是 agent 脚下的引擎。你在 <code>create_agent</code> 里用到的循环、限轮数、钩子时序，全是这张图上长出来的东西。这一篇把引擎拆开看，状态 Schema 定字段，节点干活，边定走向，Reducers 管并发写入怎么合并，这是图语言的全部词汇，学会了，中断恢复、人机协同、多智能体编排才有地基。</p>https://xiongneng.me/posts/ai/langchain/langchain-langgraph-07-memory/Wed, 07 Oct 2026 18:40:00 +0800yidao620@163.com (XiongNeng)https://xiongneng.me/posts/ai/langchain/langchain-langgraph-07-memory/AI<p><img loading="lazy" src='https://static.xiongneng.me/langchain-langgraph-cover-20261017183003.png' alt="https://static.xiongneng.me/langchain-langgraph-cover-20261017183003.png"></p> <h2 class="heading-element" id="学习进度"><span>学习进度</span> <a href="#%e5%ad%a6%e4%b9%a0%e8%bf%9b%e5%ba%a6" class="heading-mark"> <svg class="octicon octicon-link" viewBox="0 0 16 16" version="1.1" width="16" height="16" aria-hidden="true"><path d="m7.775 3.275 1.25-1.25a3.5 3.5 0 1 1 4.95 4.95l-2.5 2.5a3.5 3.5 0 0 1-4.95 0 .751.751 0 0 1 .018-1.042.751.751 0 0 1 1.042-.018 1.998 1.998 0 0 0 2.83 0l2.5-2.5a2.002 2.002 0 0 0-2.83-2.83l-1.25 1.25a.751.751 0 0 1-1.042-.018.751.751 0 0 1-.018-1.042Zm-4.69 9.64a1.998 1.998 0 0 0 2.83 0l1.25-1.25a.751.751 0 0 1 1.042.018.751.751 0 0 1 .018 1.042l-1.25 1.25a3.5 3.5 0 1 1-4.95-4.95l2.5-2.5a3.5 3.5 0 0 1 4.95 0 .751.751 0 0 1-.018 1.042.751.751 0 0 1-1.042.018 1.998 1.998 0 0 0-2.83 0l-2.5 2.5a1.998 1.998 0 0 0 0 2.83Z"></path></svg> </a> </h2><p>模型本身是无状态的，每一次调用都是初次见面，你五分钟前说过喜欢美式咖啡，五分钟后它已经不记得。这一篇给应用补上记忆，体验上才像个能长期相处的助手。</p>https://xiongneng.me/posts/ai/langchain/langchain-langgraph-06-rag/Tue, 06 Oct 2026 17:30:00 +0800yidao620@163.com (XiongNeng)https://xiongneng.me/posts/ai/langchain/langchain-langgraph-06-rag/AI<p><img loading="lazy" src='https://static.xiongneng.me/langchain-langgraph-cover-20261017175843.png' alt="https://static.xiongneng.me/langchain-langgraph-cover-20261017175843.png"></p> <h2 class="heading-element" id="学习进度"><span>学习进度</span> <a href="#%e5%ad%a6%e4%b9%a0%e8%bf%9b%e5%ba%a6" class="heading-mark"> <svg class="octicon octicon-link" viewBox="0 0 16 16" version="1.1" width="16" height="16" aria-hidden="true"><path d="m7.775 3.275 1.25-1.25a3.5 3.5 0 1 1 4.95 4.95l-2.5 2.5a3.5 3.5 0 0 1-4.95 0 .751.751 0 0 1 .018-1.042.751.751 0 0 1 1.042-.018 1.998 1.998 0 0 0 2.83 0l2.5-2.5a2.002 2.002 0 0 0-2.83-2.83l-1.25 1.25a.751.751 0 0 1-1.042-.018.751.751 0 0 1-.018-1.042Zm-4.69 9.64a1.998 1.998 0 0 0 2.83 0l1.25-1.25a.751.751 0 0 1 1.042.018.751.751 0 0 1 .018 1.042l-1.25 1.25a3.5 3.5 0 1 1-4.95-4.95l2.5-2.5a3.5 3.5 0 0 1 4.95 0 .751.751 0 0 1-.018 1.042.751.751 0 0 1-1.042.018 1.998 1.998 0 0 0-2.83 0l-2.5 2.5a1.998 1.998 0 0 0 0 2.83Z"></path></svg> </a> </h2><p>模型有两个绕不开的限制。一是上下文有限，再大的窗口也塞不下整个资料库。二是知识过期，训练数据停在某个时间点，之后的它一概不知。检索增强，RAG，就是冲着这两个限制去的，回答问题之前先把相关资料找出来给它。</p>https://xiongneng.me/posts/ai/langchain/langchain-langgraph-05-agents/Mon, 05 Oct 2026 17:05:00 +0800yidao620@163.com (XiongNeng)https://xiongneng.me/posts/ai/langchain/langchain-langgraph-05-agents/AI<p><img loading="lazy" src='https://static.xiongneng.me/langchain-langgraph-cover-20261017170947.png' alt="https://static.xiongneng.me/langchain-langgraph-cover-20261017170947.png"></p> <h2 class="heading-element" id="学习进度"><span>学习进度</span> <a href="#%e5%ad%a6%e4%b9%a0%e8%bf%9b%e5%ba%a6" class="heading-mark"> <svg class="octicon octicon-link" viewBox="0 0 16 16" version="1.1" width="16" height="16" aria-hidden="true"><path d="m7.775 3.275 1.25-1.25a3.5 3.5 0 1 1 4.95 4.95l-2.5 2.5a3.5 3.5 0 0 1-4.95 0 .751.751 0 0 1 .018-1.042.751.751 0 0 1 1.042-.018 1.998 1.998 0 0 0 2.83 0l2.5-2.5a2.002 2.002 0 0 0-2.83-2.83l-1.25 1.25a.751.751 0 0 1-1.042-.018.751.751 0 0 1-.018-1.042Zm-4.69 9.64a1.998 1.998 0 0 0 2.83 0l1.25-1.25a.751.751 0 0 1 1.042.018.751.751 0 0 1 .018 1.042l-1.25 1.25a3.5 3.5 0 1 1-4.95-4.95l2.5-2.5a3.5 3.5 0 0 1 4.95 0 .751.751 0 0 1-.018 1.042.751.751 0 0 1-1.042.018 1.998 1.998 0 0 0-2.83 0l-2.5 2.5a1.998 1.998 0 0 0 0 2.83Z"></path></svg> </a> </h2><p>Agent 是一台完整机器，模型是发动机，<code>create_agent</code> 是把油路电路接好的线束，系统提示、工具、中间件都在里面。这一篇把 create_agent 的完整签名摊开看，十五个参数逐个过，六个钩子各守一个位置跑通，流式执行也顺一遍。</p>https://xiongneng.me/posts/ai/langchain/langchain-langgraph-04-tools/Sun, 04 Oct 2026 16:15:00 +0800yidao620@163.com (XiongNeng)https://xiongneng.me/posts/ai/langchain/langchain-langgraph-04-tools/AI<p><img loading="lazy" src='https://static.xiongneng.me/langchain-langgraph-cover-20261017161835.png' alt="https://static.xiongneng.me/langchain-langgraph-cover-20261017161835.png"></p> <h2 class="heading-element" id="学习进度"><span>学习进度</span> <a href="#%e5%ad%a6%e4%b9%a0%e8%bf%9b%e5%ba%a6" class="heading-mark"> <svg class="octicon octicon-link" viewBox="0 0 16 16" version="1.1" width="16" height="16" aria-hidden="true"><path d="m7.775 3.275 1.25-1.25a3.5 3.5 0 1 1 4.95 4.95l-2.5 2.5a3.5 3.5 0 0 1-4.95 0 .751.751 0 0 1 .018-1.042.751.751 0 0 1 1.042-.018 1.998 1.998 0 0 0 2.83 0l2.5-2.5a2.002 2.002 0 0 0-2.83-2.83l-1.25 1.25a.751.751 0 0 1-1.042-.018.751.751 0 0 1-.018-1.042Zm-4.69 9.64a1.998 1.998 0 0 0 2.83 0l1.25-1.25a.751.751 0 0 1 1.042.018.751.751 0 0 1 .018 1.042l-1.25 1.25a3.5 3.5 0 1 1-4.95-4.95l2.5-2.5a3.5 3.5 0 0 1 4.95 0 .751.751 0 0 1-.018 1.042.751.751 0 0 1-1.042.018 1.998 1.998 0 0 0-2.83 0l-2.5 2.5a1.998 1.998 0 0 0 0 2.83Z"></path></svg> </a> </h2><p>工具让模型伸手干活。你给它的工具箱里放一个 <code>get_weather</code>，它自己决定调不调、自己填参数，拿到结果再组织回答。这一篇把工具这条链路补齐，从 <code>@tool</code> 装饰器把一个普通函数变成工具讲起，到 bind_tools 绑定、模型发起调用、ToolMessage 回传，再到出错怎么办、工具多了怎么管。</p>https://xiongneng.me/posts/ai/langchain/langchain-langgraph-03-messages-structured-output/Sat, 03 Oct 2026 15:20:00 +0800yidao620@163.com (XiongNeng)https://xiongneng.me/posts/ai/langchain/langchain-langgraph-03-messages-structured-output/AI<p><img loading="lazy" src='https://static.xiongneng.me/langchain-langgraph-cover-20261017150727.png' alt="https://static.xiongneng.me/langchain-langgraph-cover-20261017150727.png"></p> <h2 class="heading-element" id="学习进度"><span>学习进度</span> <a href="#%e5%ad%a6%e4%b9%a0%e8%bf%9b%e5%ba%a6" class="heading-mark"> <svg class="octicon octicon-link" viewBox="0 0 16 16" version="1.1" width="16" height="16" aria-hidden="true"><path d="m7.775 3.275 1.25-1.25a3.5 3.5 0 1 1 4.95 4.95l-2.5 2.5a3.5 3.5 0 0 1-4.95 0 .751.751 0 0 1 .018-1.042.751.751 0 0 1 1.042-.018 1.998 1.998 0 0 0 2.83 0l2.5-2.5a2.002 2.002 0 0 0-2.83-2.83l-1.25 1.25a.751.751 0 0 1-1.042-.018.751.751 0 0 1-.018-1.042Zm-4.69 9.64a1.998 1.998 0 0 0 2.83 0l1.25-1.25a.751.751 0 0 1 1.042.018.751.751 0 0 1 .018 1.042l-1.25 1.25a3.5 3.5 0 1 1-4.95-4.95l2.5-2.5a3.5 3.5 0 0 1 4.95 0 .751.751 0 0 1-.018 1.042.751.751 0 0 1-1.042.018 1.998 1.998 0 0 0-2.83 0l-2.5 2.5a1.998 1.998 0 0 0 0 2.83Z"></path></svg> </a> </h2><p>消息是模型接口的通用语言。你传进 <code>invoke</code> 的一切，最后都会被整理成一个消息列表发给模型，字符串只是框架帮你包好的语法糖。多轮对话、多模态、工具调用，这几个场景里你必须亲手构造消息，所以先把这门语言讲透。</p>https://xiongneng.me/posts/ai/langchain/langchain-langgraph-02-models/Fri, 02 Oct 2026 13:10:00 +0800yidao620@163.com (XiongNeng)https://xiongneng.me/posts/ai/langchain/langchain-langgraph-02-models/AI<p><img loading="lazy" src='https://static.xiongneng.me/langchain-langgraph-cover-20261017122621.png' alt="https://static.xiongneng.me/langchain-langgraph-cover-20261017122621.png"></p> <h2 class="heading-element" id="学习进度"><span>学习进度</span> <a href="#%e5%ad%a6%e4%b9%a0%e8%bf%9b%e5%ba%a6" class="heading-mark"> <svg class="octicon octicon-link" viewBox="0 0 16 16" version="1.1" width="16" height="16" aria-hidden="true"><path d="m7.775 3.275 1.25-1.25a3.5 3.5 0 1 1 4.95 4.95l-2.5 2.5a3.5 3.5 0 0 1-4.95 0 .751.751 0 0 1 .018-1.042.751.751 0 0 1 1.042-.018 1.998 1.998 0 0 0 2.83 0l2.5-2.5a2.002 2.002 0 0 0-2.83-2.83l-1.25 1.25a.751.751 0 0 1-1.042-.018.751.751 0 0 1-.018-1.042Zm-4.69 9.64a1.998 1.998 0 0 0 2.83 0l1.25-1.25a.751.751 0 0 1 1.042.018.751.751 0 0 1 .018 1.042l-1.25 1.25a3.5 3.5 0 1 1-4.95-4.95l2.5-2.5a3.5 3.5 0 0 1 4.95 0 .751.751 0 0 1-.018 1.042.751.751 0 0 1-1.042.018 1.998 1.998 0 0 0-2.83 0l-2.5 2.5a1.998 1.998 0 0 0 0 2.83Z"></path></svg> </a> </h2><p>模型是应用的大脑，这一篇讲怎么把它接进来。LangChain 的模型抽象把所有提供商收拢到一个接口后面，你的代码只面对 <code>BaseChatModel</code>，面对 <code>invoke</code>、<code>stream</code>、<code>batch</code> 三个方法，换模型只改一个字符串，不用再碰每家 SDK 各自的鉴权、参数名和流式协议。</p>https://xiongneng.me/posts/ai/langchain/langchain-langgraph-01-init/Thu, 01 Oct 2026 09:50:00 +0800yidao620@163.com (XiongNeng)https://xiongneng.me/posts/ai/langchain/langchain-langgraph-01-init/AI<p>现在做 AI 应用的人，几乎都绕不开 LangChain 和 LangGraph 这两个名字。</p> <p>大模型本身只会聊天，你问一句它答一句。要把这种能力变成能干活的应用，还得给它接上工具，查数据库、调接口、读文件，还要记住多轮对话里聊过什么。流程长一点，需要能在中间停下来等人确认；跑上十分钟的任务中途挂了，得从断点续上，而不是从头再来。这些事情每件都不难，凑在一起就是一套基础设施。</p>