07-08-周三_10-09-42

14_LangGraph概述.ipynb

@@ -0,0 +1,299 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# 14 LangGraph 概述\n",
+    "\n",
+    "## 学习目标\n",
+    "1. 理解为什么需要 LangGraph：复杂智能体流程的控制需求\n",
+    "2. 掌握 LangGraph 的核心概念：StateGraph、State、Node、Edge\n",
+    "3. 理解 LangGraph 与 LangChain 的关系和分工\n",
+    "4. 能够编写一个最简单的 LangGraph 状态图\n",
+    "5. 了解 LangGraph 的典型应用场景"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 1. 为什么需要 LangGraph\n",
+    "\n",
+    "前面我们用 LangChain 构建了链式应用和简单的工具调用智能体。但当智能体流程变得复杂时，会遇到一些问题：\n",
+    "\n",
+    "- **流程不可控**：标准 Agent 的黑盒决策难以调试和约束\n",
+    "- **循环和分支困难**：难以实现多轮循环、条件分支、人机协作\n",
+    "- **状态管理复杂**：多步骤之间的状态传递不清晰\n",
+    "- **缺乏持久化**：无法暂停、恢复或检查执行历史\n",
+    "\n",
+    "**LangGraph** 就是为了解决这些问题而设计的。它把智能体流程建模为一个**状态图（StateGraph）**，让开发者精确控制每个步骤。"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 2. LangGraph 是什么\n",
+    "\n",
+    "LangGraph 是 LangChain 生态中的一个库，用于构建基于**状态图**的智能体应用。它的核心思想是：\n",
+    "\n",
+    "- 把应用看作一个图（Graph）\n",
+    "- 图中的每个节点（Node）是一个函数或操作\n",
+    "- 边（Edge）控制流程走向\n",
+    "- 状态（State）在节点之间传递\n",
+    "\n",
+    "LangGraph 基于 LangChain 的 Runnable 接口构建，因此可以无缝使用 LangChain 的模型、工具、Prompt 等组件。"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 3. LangGraph 与 LangChain 的关系\n",
+    "\n",
+    "| 对比项 | LangChain | LangGraph |\n",
+    "| --- | --- | --- |\n",
+    "| 核心抽象 | Chain / Agent | StateGraph |\n",
+    "| 流程控制 | 线性或简单循环 | 任意图结构、条件分支、循环 |\n",
+    "| 状态管理 | 隐式传递 | 显式 State 对象 |\n",
+    "| 适用场景 | 简单流水线、单次调用 | 复杂多智能体、多轮交互 |\n",
+    "| 关系 | 基础组件库 | 基于 LangChain 构建的流程编排层 |\n",
+    "\n",
+    "可以把 LangChain 理解为「零件库」，LangGraph 理解为「装配线控制系统」。"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 4. 核心概念\n",
+    "\n",
+    "### 4.1 State（状态）\n",
+    "状态是整个图共享的数据结构。每个节点读取状态、修改状态，然后把状态传递给下一个节点。\n",
+    "\n",
+    "### 4.2 StateGraph（状态图）\n",
+    "状态图是 LangGraph 最核心的类。它负责定义节点、边和状态类型。\n",
+    "\n",
+    "### 4.3 Node（节点）\n",
+    "节点是图中的执行单元，通常是一个 Python 函数。节点接收当前状态，返回对状态的更新。\n",
+    "\n",
+    "### 4.4 Edge（边）\n",
+    "边定义节点之间的连接关系。LangGraph 支持：\n",
+    "- 普通边：固定从一个节点到另一个节点\n",
+    "- 条件边：根据状态决定下一个节点\n",
+    "- START / END：图的起点和终点"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 5. 第一个 LangGraph 程序\n",
+    "\n",
+    "下面构建一个最简单的图：两个节点 `node_a` 和 `node_b`，`node_a` 给状态加 1，`node_b` 再给状态乘 2。"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from typing_extensions import TypedDict\n",
+    "from langgraph.graph import StateGraph, START, END\n",
+    "\n",
+    "# 1. 定义状态类型\n",
+    "class MyState(TypedDict):\n",
+    "    \"\"\"图中的共享状态。\"\"\"\n",
+    "    value: int\n",
+    "\n",
+    "# 2. 定义节点函数\n",
+    "def node_a(state: MyState):\n",
+    "    \"\"\"把 value 加 1。\"\"\"\n",
+    "    print(f'Node A 接收：{state[\"value\"]}')\n",
+    "    return {'value': state['value'] + 1}\n",
+    "\n",
+    "def node_b(state: MyState):\n",
+    "    \"\"\"把 value 乘 2。\"\"\"\n",
+    "    print(f'Node B 接收：{state[\"value\"]}')\n",
+    "    return {'value': state['value'] * 2}\n",
+    "\n",
+    "# 3. 构建状态图\n",
+    "builder = StateGraph(MyState)\n",
+    "builder.add_node('node_a', node_a)\n",
+    "builder.add_node('node_b', node_b)\n",
+    "\n",
+    "# 4. 添加边\n",
+    "builder.add_edge(START, 'node_a')\n",
+    "builder.add_edge('node_a', 'node_b')\n",
+    "builder.add_edge('node_b', END)\n",
+    "\n",
+    "# 5. 编译图\n",
+    "graph = builder.compile()\n",
+    "\n",
+    "# 6. 运行图\n",
+    "result = graph.invoke({'value': 3})\n",
+    "print(f'\\n最终结果：{result}')"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### 代码解释\n",
+    "\n",
+    "- `TypedDict` 定义状态结构，告诉 LangGraph 图中有哪些字段\n",
+    "- `StateGraph(MyState)` 创建状态图实例\n",
+    "- `add_node(name, func)` 添加节点\n",
+    "- `add_edge(START, 'node_a')` 设置从图起点到 node_a 的边\n",
+    "- `compile()` 编译图，生成可执行的 Runnable\n",
+    "- `invoke({'value': 3})` 运行图，传入初始状态\n",
+    "\n",
+    "这个例子的执行流程是：`START -> node_a(value=3 -> 4) -> node_b(value=4 -> 8) -> END`，最终结果是 8。"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 6. LangGraph 执行流程可视化\n",
+    "\n",
+    "LangGraph 提供了 `get_graph()` 方法，可以查看图的结构。\n",
+    "\n",
+    "在线查看Mermaid 图：https://mermaid.live/edit\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# 打印图的节点和边\n",
+    "print(graph.get_graph().nodes)\n",
+    "print(graph.get_graph().edges)\n",
+    "\n",
+    "# 使用 mermaid 格式可视化（在支持 mermaid 的 Markdown 查看器中可显示）\n",
+    "print('\\nMermaid 图：')\n",
+    "print(graph.get_graph().draw_mermaid())"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 7. 在 LangGraph 中使用 LangChain 组件\n",
+    "\n",
+    "LangGraph 的节点可以使用任何 LangChain 组件。下面是一个使用 LLM 和 Prompt 的节点示例。"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from langchain_openai import ChatOpenAI\n",
+    "from langchain_core.prompts import ChatPromptTemplate\n",
+    "from langgraph.graph.message import add_messages\n",
+    "from dotenv import load_dotenv\n",
+    "\n",
+    "load_dotenv()\n",
+    "\n",
+    "\n",
+    "# 定义带消息历史的状态\n",
+    "class ChatState(TypedDict):\n",
+    "    messages: list  # 对话历史\n",
+    "\n",
+    "llm = ChatOpenAI(model='qwen3.6-35b-A3b', temperature=0.7)\n",
+    "\n",
+    "def chat_node(state: ChatState):\n",
+    "    \"\"\"调用 LLM 回复用户。\"\"\"\n",
+    "    prompt = ChatPromptTemplate.from_messages([\n",
+    "        ('system', '你是友好的 AI 助手。'),\n",
+    "        *state['messages']\n",
+    "    ])\n",
+    "    response = (prompt | llm).invoke({})\n",
+    "    return {'messages': [response]}\n",
+    "\n",
+    "builder = StateGraph(ChatState)\n",
+    "builder.add_node('chat', chat_node)\n",
+    "builder.add_edge(START, 'chat')\n",
+    "builder.add_edge('chat', END)\n",
+    "\n",
+    "chat_graph = builder.compile()\n",
+    "\n",
+    "result = chat_graph.invoke({\n",
+    "    'messages': [('user', '请用一句话介绍 LangGraph')]\n",
+    "})\n",
+    "print(result['messages'][-1].content)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### 代码解释\n",
+    "\n",
+    "- `messages` 是列表类型，每个元素是 `(role, content)` 元组\n",
+    "- `ChatPromptTemplate.from_messages([...])` 中的 `*state['messages']` 把历史消息展开\n",
+    "- 节点返回 `{'messages': [response]}`，LangGraph 会自动合并到状态中\n",
+    "- 这里只是一个单节点图，后续课程会扩展到多轮对话和循环"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 8. LangGraph 的典型应用场景\n",
+    "\n",
+    "LangGraph 特别适合以下场景：\n",
+    "\n",
+    "| 场景 | 说明 |\n",
+    "| --- | --- |\n",
+    "| **多轮工具调用** | 在循环中反复调用工具，直到获得足够信息 |\n",
+    "| **人机协作（Human-in-the-loop）** | 在关键节点暂停，等待人类确认 |\n",
+    "| **多智能体系统** | 多个 Agent 通过图结构协作完成任务 |\n",
+    "| **复杂审批流程** | 根据条件分支决定流程走向 |\n",
+    "| **长期记忆** | 通过持久化状态实现跨会话记忆 |\n",
+    "\n",
+    "这些场景在后面几节课会逐一展开。"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 9. 本节课练习\n",
+    "\n",
+    "1. 修改第一个 LangGraph 示例，增加第三个节点 `node_c`，实现 value = value - 2\n",
+    "2. 改变三个节点的连接顺序，观察最终结果如何变化\n",
+    "3. 在 ChatState 中增加一个 `topic` 字段，让 system prompt 根据 topic 动态变化\n",
+    "4. 尝试画出本节课第一个示例的状态图（START -> node_a -> node_b -> END）\n",
+    "5. 思考：LangGraph 与你之前学过的 LangChain Chain 相比，最大的优势是什么？"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": ".venv",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.14.4"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 4
+}