07-08-周三_10-09-42

06_Prompt模板.ipynb

@@ -0,0 +1,579 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# 06 Prompt 模板\n",
+    "\n",
+    "## 学习目标\n",
+    "1. 理解 Prompt 模板的作用：将 Prompt 从硬编码变为可复用、可配置的组件\n",
+    "2. 掌握两种最常用的模板类型：PromptTemplate 和 ChatPromptTemplate\n",
+    "3. 学会使用变量占位符 {name} 动态填充内容\n",
+    "4. 理解 Few-Shot Prompt 少样本提示的构建方法\n",
+    "5. 了解常见的 Prompt Engineering 技巧及其对模型输出的影响"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 1. 为什么需要 Prompt 模板\n",
+    "\n",
+    "直接拼接字符串来构造 Prompt 会带来几个问题：\n",
+    "\n",
+    "- **可读性差**：Python 字符串拼接难以看清最终给模型的提示长什么样\n",
+    "- **容易出错**：变量多的时候容易遗漏、顺序错乱\n",
+    "- **难以复用**：同样的提示结构无法直接复用到不同输入\n",
+    "- **难以管理**：无法集中管理和版本化 Prompt\n",
+    "\n",
+    "**Prompt 模板** 就是解决这些问题的方案。它允许你：\n",
+    "\n",
+    "- 把 Prompt 写成一个带占位符的模板\n",
+    "- 运行时传入变量自动填充\n",
+    "- 复用同一个结构处理不同输入\n",
+    "- 把提示逻辑与业务逻辑分离"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 2. PromptTemplate：基础字符串模板\n",
+    "\n",
+    "`PromptTemplate` 是最基础的模板类，适用于构造单条文本提示。它只有一个 `template` 字符串和若干 {变量}。"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from langchain_openai import ChatOpenAI\n",
+    "from langchain_core.prompts import ChatPromptTemplate\n",
+    "from dotenv import load_dotenv\n",
+    "\n",
+    "load_dotenv()\n",
+    "\n",
+    "\n",
+    "# 创建模型\n",
+    "llm = ChatOpenAI(model='qwen3.6-35b-A3b', temperature=0.7)\n",
+    "\n",
+    "# 使用元组列表定义消息模板\n",
+    "prompt = ChatPromptTemplate.from_messages([\n",
+    "    ('system', '你是一位{role}，擅长用{style}风格回答问题。'),\n",
+    "    ('user', '{question}')\n",
+    "])\n",
+    "\n",
+    "# 查看组装后的 Prompt\n",
+    "input_vars = {\n",
+    "    'role': '计算机科学家',\n",
+    "    'style': '通俗易懂',\n",
+    "    'question': '什么是递归？'\n",
+    "}\n",
+    "print('===== 组装后的 Prompt =====')\n",
+    "for msg in prompt.format_messages(**input_vars):\n",
+    "    print(f'{msg.type}: {msg.content}')\n",
+    "\n",
+    "# 构建链并调用\n",
+    "chain = prompt | llm\n",
+    "print('\\n===== 大模型输出 =====')\n",
+    "result = chain.invoke(input_vars)\n",
+    "print(result.content)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### 2.1 使用 from_template 快速创建\n",
+    "\n",
+    "如果你习惯简单写法，可以使用 `from_template` 类方法，自动识别模板中的变量。"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from langchain_openai import ChatOpenAI\n",
+    "from langchain_core.prompts import ChatPromptTemplate\n",
+    "from langchain_core.prompts import SystemMessagePromptTemplate, HumanMessagePromptTemplate\n",
+    "\n",
+    "# 创建模型\n",
+    "llm = ChatOpenAI(model='qwen3.6-35b-A3b', temperature=0.7)\n",
+    "\n",
+    "prompt = ChatPromptTemplate.from_messages([\n",
+    "    SystemMessagePromptTemplate.from_template('你是一位耐心的老师。'),\n",
+    "    HumanMessagePromptTemplate.from_template('请解释{concept}，要求{requirement}。')\n",
+    "])\n",
+    "\n",
+    "# 查看组装后的 Prompt\n",
+    "input_vars = {\n",
+    "    'concept': '神经网络',\n",
+    "    'requirement': '用小学生能听懂的语言'\n",
+    "}\n",
+    "print('===== 组装后的 Prompt =====')\n",
+    "for msg in prompt.format_messages(**input_vars):\n",
+    "    print(f'{msg.type}: {msg.content}')\n",
+    "\n",
+    "# 构建链并调用\n",
+    "chain = prompt | llm\n",
+    "print('\\n===== 大模型输出 =====')\n",
+    "result = chain.invoke(input_vars)\n",
+    "print(result.content)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 3. ChatPromptTemplate：聊天消息模板\n",
+    "\n",
+    "现代大模型通常采用对话格式，每条消息都有角色（role）。`ChatPromptTemplate` 就是为这种场景设计的。\n",
+    "\n",
+    "常见的消息类型：\n",
+    "\n",
+    "| 消息类型 | 作用 |\n",
+    "| --- | --- |\n",
+    "| system | 设定模型的身份、能力和行为规则 |\n",
+    "| user / human | 用户的输入问题 |\n",
+    "| assistant / ai | 模型的回复，可用于 Few-Shot 示例 |\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from langchain_openai import ChatOpenAI\n",
+    "from langchain_core.prompts import ChatPromptTemplate\n",
+    "\n",
+    "# 创建模型\n",
+    "llm = ChatOpenAI(model='qwen3.6-35b-A3b', temperature=0.7)\n",
+    "\n",
+    "prompt = ChatPromptTemplate.from_messages([\n",
+    "    ('system', '你是一位{role}。'),\n",
+    "    ('user', '请{action}：{content}')\n",
+    "])\n",
+    "\n",
+    "# 先固定 role 变量，得到一个新的模板\n",
+    "translator_prompt = prompt.partial(role='专业翻译')\n",
+    "\n",
+    "# 查看组装后的 Prompt\n",
+    "input_vars = {\n",
+    "    'action': '把以下中文翻译成英文',\n",
+    "    'content': '今天天气很好'\n",
+    "}\n",
+    "print('===== 组装后的 Prompt =====')\n",
+    "for msg in translator_prompt.format_messages(**input_vars):\n",
+    "    print(f'{msg.type}: {msg.content}')\n",
+    "\n",
+    "# 构建链并调用\n",
+    "chain = translator_prompt | llm\n",
+    "print('\\n===== 大模型输出 =====')\n",
+    "result = chain.invoke(input_vars)\n",
+    "print(result.content)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### 3.1 使用消息模板对象构建模板\n",
+    "\n",
+    "除了元组，也可以使用 LangChain 提供的**消息模板类**来构建模板。这种方式语义更清晰，而且同样支持变量替换。\n",
+    "\n",
+    "**注意区分两类对象**：\n",
+    "- `SystemMessage` / `HumanMessage`：普通消息对象，内容固定，不会解析 `{变量}`\n",
+    "- `SystemMessagePromptTemplate` / `HumanMessagePromptTemplate`：消息模板对象，内容中的 `{变量}` 会被替换"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from langchain_openai import ChatOpenAI\n",
+    "from langchain_core.prompts import ChatPromptTemplate\n",
+    "\n",
+    "# 创建模型\n",
+    "llm = ChatOpenAI(model='qwen3.6-35b-A3b', temperature=0.1)\n",
+    "\n",
+    "prompt = ChatPromptTemplate.from_messages([\n",
+    "    ('system', '你是一个情绪分析助手，只输出正面或负面。'),\n",
+    "    ('human', '这个产品太差了'),\n",
+    "    ('ai', '负面'),\n",
+    "    ('human', '这次的体验非常愉快'),\n",
+    "    ('ai', '正面'),\n",
+    "    ('human', '{text}')\n",
+    "])\n",
+    "\n",
+    "# 查看组装后的 Prompt\n",
+    "input_vars = {'text': '物流速度还可以，但包装破损了'}\n",
+    "print('===== 组装后的 Prompt =====')\n",
+    "for msg in prompt.format_messages(**input_vars):\n",
+    "    print(f'{msg.type}: {msg.content}')\n",
+    "\n",
+    "# 构建链并调用\n",
+    "chain = prompt | llm\n",
+    "print('\\n===== 大模型输出 =====')\n",
+    "result = chain.invoke(input_vars)\n",
+    "print(result.content)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 4. 模板变量的高级用法\n",
+    "\n",
+    "### 4.1 默认值与 partial\n",
+    "\n",
+    "你可以先填充部分变量，得到一个新的模板，后续再填充剩余变量。"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from langchain_openai import ChatOpenAI\n",
+    "from langchain_core.prompts import ChatPromptTemplate, FewShotChatMessagePromptTemplate\n",
+    "\n",
+    "# 创建模型\n",
+    "llm = ChatOpenAI(model='qwen3.6-35b-A3b', temperature=0.1)\n",
+    "\n",
+    "# 准备示例数据\n",
+    "examples = [\n",
+    "    {'input': '苹果', 'output': 'apple'},\n",
+    "    {'input': '香蕉', 'output': 'banana'},\n",
+    "    {'input': '人工智能', 'output': 'artificial intelligence'},\n",
+    "]\n",
+    "\n",
+    "# 定义每个示例的格式\n",
+    "example_prompt = ChatPromptTemplate.from_messages([\n",
+    "    ('human', '{input}'),\n",
+    "    ('ai', '{output}')\n",
+    "])\n",
+    "\n",
+    "# 构建 Few-Shot 提示模板\n",
+    "few_shot_prompt = FewShotChatMessagePromptTemplate(\n",
+    "    example_prompt=example_prompt,\n",
+    "    examples=examples\n",
+    ")\n",
+    "\n",
+    "# 组合成完整提示\n",
+    "final_prompt = ChatPromptTemplate.from_messages([\n",
+    "    ('system', '你是一个中英文翻译助手，请参考示例回答问题。'),\n",
+    "    few_shot_prompt,\n",
+    "    ('human', '{input}')\n",
+    "])\n",
+    "\n",
+    "# 查看组装后的 Prompt\n",
+    "input_vars = {'input': '机器学习'}\n",
+    "print('===== 组装后的 Prompt =====')\n",
+    "for msg in final_prompt.format_messages(**input_vars):\n",
+    "    print(f'{msg.type}: {msg.content}')\n",
+    "\n",
+    "# 构建链并调用\n",
+    "chain = final_prompt | llm\n",
+    "print('\\n===== 大模型输出 =====')\n",
+    "result = chain.invoke(input_vars)\n",
+    "print(result.content)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### 4.2 多轮对话模板\n",
+    "\n",
+    "模板中可以预设多轮对话历史，常用于构建 Few-Shot 示例或保持上下文："
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from langchain_core.prompts import ChatPromptTemplate\n",
+    "\n",
+    "prompt = ChatPromptTemplate.from_messages([\n",
+    "    ('system', '你是一个情绪分析助手，只输出正面或负面。'),\n",
+    "    ('human', '这个产品太差了'),\n",
+    "    ('ai', '负面'),\n",
+    "    ('human', '这次的体验非常愉快'),\n",
+    "    ('ai', '正面'),\n",
+    "    ('human', '{text}')\n",
+    "])\n",
+    "\n",
+    "messages = prompt.format_messages(text='物流速度还可以，但包装破损了')\n",
+    "for msg in messages:\n",
+    "    print(f'{msg.type}: {msg.content}')"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 5. Few-Shot Prompt：少样本提示\n",
+    "\n",
+    "Few-Shot Prompt 是 Prompt Engineering 中最有效的技巧之一。它通过在输入前给出若干「示例-答案」对，让模型学习期望的输出格式和风格。\n",
+    "\n",
+    "LangChain 提供了 `FewShotPromptTemplate` 和 `FewShotChatMessagePromptTemplate` 两种主要方式。"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from langchain_core.prompts import ChatPromptTemplate, FewShotChatMessagePromptTemplate\n",
+    "\n",
+    "# 准备示例数据\n",
+    "examples = [\n",
+    "    {'input': '苹果', 'output': 'apple'},\n",
+    "    {'input': '香蕉', 'output': 'banana'},\n",
+    "    {'input': '人工智能', 'output': 'artificial intelligence'},\n",
+    "]\n",
+    "\n",
+    "# 定义每个示例的格式\n",
+    "example_prompt = ChatPromptTemplate.from_messages([\n",
+    "    ('human', '{input}'),\n",
+    "    ('ai', '{output}')\n",
+    "])\n",
+    "\n",
+    "# 构建 Few-Shot 提示模板\n",
+    "few_shot_prompt = FewShotChatMessagePromptTemplate(\n",
+    "    example_prompt=example_prompt,\n",
+    "    examples=examples\n",
+    ")\n",
+    "\n",
+    "# 组合成完整提示\n",
+    "final_prompt = ChatPromptTemplate.from_messages([\n",
+    "    ('system', '你是一个中英文翻译助手，请参考示例回答问题。'),\n",
+    "    few_shot_prompt,\n",
+    "    ('human', '{input}')\n",
+    "])\n",
+    "\n",
+    "messages = final_prompt.format_messages(input='机器学习')\n",
+    "for msg in messages:\n",
+    "    print(f'{msg.type}: {msg.content}')"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 6. Prompt Engineering 技巧\n",
+    "\n",
+    "Prompt Engineering 是指通过设计和优化提示词，让模型输出更准确、更符合预期。下面介绍几个最常用、最有效的技巧。"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### 6.1 角色扮演（Role Prompting）\n",
+    "\n",
+    "给模型设定一个清晰的角色，可以显著提升回答质量。"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from langchain_openai import ChatOpenAI\n",
+    "from langchain_core.prompts import ChatPromptTemplate\n",
+    "\n",
+    "llm = ChatOpenAI(model='qwen3.6-35b-A3b', temperature=0.7)\n",
+    "\n",
+    "# 无角色设定\n",
+    "prompt1 = ChatPromptTemplate.from_messages([\n",
+    "    ('user', '解释什么是区块链')\n",
+    "])\n",
+    "\n",
+    "# 有角色设定\n",
+    "prompt2 = ChatPromptTemplate.from_messages([\n",
+    "    ('system', '你是一位资深技术讲师，擅长用比喻和例子解释复杂概念。'),\n",
+    "    ('user', '解释什么是区块链')\n",
+    "])\n",
+    "\n",
+    "chain1 = prompt1 | llm\n",
+    "chain2 = prompt2 | llm\n",
+    "\n",
+    "print('=== 无角色 ===')\n",
+    "print(chain1.invoke({}).content[:250])\n",
+    "print('\\n=== 有角色 ===')\n",
+    "print(chain2.invoke({}).content[:250])"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### 6.2 明确输出格式（Output Formatting）\n",
+    "\n",
+    "如果你希望模型按特定格式输出，要在 Prompt 中明确说明。"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from langchain_openai import ChatOpenAI\n",
+    "from langchain_core.prompts import ChatPromptTemplate\n",
+    "from langchain_core.output_parsers import JsonOutputParser\n",
+    "\n",
+    "llm = ChatOpenAI(model='qwen3.6-35b-A3b', temperature=0.7)\n",
+    "\n",
+    "# 把 JSON 格式示例放在 system 消息中\n",
+    "# 注意：{ 和 } 是 LangChain 的变量占位符，如果要显示字面量大括号，必须写成 {{ 和 }}\n",
+    "system_template = '''你是一个信息提取助手。请只输出 JSON 格式，不要包含任何解释。\n",
+    "\n",
+    "输出格式如下：\n",
+    "{{\n",
+    "  \"names\": [...],\n",
+    "  \"locations\": [...],\n",
+    "  \"time\": \"\"\n",
+    "}}'''\n",
+    "\n",
+    "prompt = ChatPromptTemplate.from_messages([\n",
+    "    ('system', system_template),\n",
+    "    ('user', '文本：{text}')\n",
+    "])\n",
+    "\n",
+    "chain = prompt | llm | JsonOutputParser()\n",
+    "\n",
+    "input_vars = {\n",
+    "    'text': '2024年5月1日，李明和王芳一起去了北京故宫参观。'\n",
+    "}\n",
+    "\n",
+    "print('===== 组装后的 Prompt =====')\n",
+    "for msg in prompt.format_messages(**input_vars):\n",
+    "    print(f'{msg.type}: {msg.content}')\n",
+    "\n",
+    "print('\\n===== 大模型输出 =====')\n",
+    "result = chain.invoke(input_vars)\n",
+    "print(result)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### 6.3 分步骤思考（Chain-of-Thought）\n",
+    "\n",
+    "对于推理类问题，让模型「一步步思考」往往能显著提升准确率。"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from langchain_openai import ChatOpenAI\n",
+    "from langchain_core.prompts import ChatPromptTemplate\n",
+    "\n",
+    "llm = ChatOpenAI(model='qwen3.6-35b-A3b', temperature=0.1)\n",
+    "\n",
+    "# 直接提问\n",
+    "prompt1 = ChatPromptTemplate.from_messages([\n",
+    "    ('user', '问题：一个农场有鸡和兔共35只，脚共94只。鸡和兔各几只？请直接给出答案。')\n",
+    "])\n",
+    "\n",
+    "# 分步思考\n",
+    "prompt2 = ChatPromptTemplate.from_messages([\n",
+    "    ('user', '问题：一个农场有鸡和兔共35只，脚共94只。鸡和兔各几只？请一步步思考并给出答案。')\n",
+    "])\n",
+    "\n",
+    "print('=== 直接提问 ===')\n",
+    "print((prompt1 | llm).invoke({}).content)\n",
+    "print('\\n=== 分步思考 ===')\n",
+    "print((prompt2 | llm).invoke({}).content)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 7. 完整示例：Prompt 模板与链结合\n",
+    "\n",
+    "下面是一个综合示例：使用模板构建一个「学习助手」，根据学生年级和知识点生成适合难度的解释。"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from langchain_openai import ChatOpenAI\n",
+    "from langchain_core.prompts import ChatPromptTemplate\n",
+    "from langchain_core.output_parsers import StrOutputParser\n",
+    "\n",
+    "llm = ChatOpenAI(model='qwen3.6-35b-A3b', temperature=0.7)\n",
+    "\n",
+    "prompt = ChatPromptTemplate.from_messages([\n",
+    "    ('system', '你是一位{grade}水平的优秀教师。'),\n",
+    "    ('user', '请用适合{grade}学生理解的方式解释 {topic}，并举一个生活中的例子。')\n",
+    "])\n",
+    "\n",
+    "chain = prompt | llm | StrOutputParser()\n",
+    "\n",
+    "result = chain.invoke({\n",
+    "    'grade': '小学三年级',\n",
+    "    'topic': '浮力'\n",
+    "})\n",
+    "print(result)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 8. 本节课练习\n",
+    "\n",
+    "1. 使用 `PromptTemplate` 创建一个邮件生成模板，变量包含收件人姓名、主题、正文要点，调用 `format` 输出完整邮件\n",
+    "2. 使用 `ChatPromptTemplate` 创建一个「代码审查助手」模板，system 设定角色，user 传入代码片段\n",
+    "3. 使用 `.partial()` 固定 system 角色为「技术文档写手」，只传入 user 变量运行\n",
+    "4. 使用 `FewShotChatMessagePromptTemplate` 创建一个三示例的情绪分类器，然后输入新句子测试\n",
+    "5. 对比实验：分别用「直接提问」和「分步骤思考」两种方式向模型提问同一道数学题，观察输出差异"
+   ]
+  }
+ ],
+ "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": "3.11.3"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 4
+}