工作流和 Agent¶

本指南回顾了 Agent 系统的常见模式。在描述这些系统时，区分“工作流”和“Agent”可能很有用。Anthropic 在这里很好地解释了这种区别的一种思考方式

工作流是通过预定义的代码路径编排 LLM 和工具的系统。另一方面，Agent 是 LLM 动态指导自身流程和工具使用的系统，保持对其完成任务方式的控制。

这是一个可视化这些差异的简单方法

Agent Workflow

在构建 Agent 和工作流时，LangGraph 提供了许多优势，包括持久性、流式传输以及对调试和部署的支持。

设置¶

注意

Functional API 需要 @langchain/langgraph>=0.2.24。

您可以使用任何支持结构化输出和工具调用的聊天模型。下面，我们展示了安装软件包、设置 API 密钥以及测试 Anthropic 的结构化输出/工具调用的过程。

安装依赖项

yarn add @langchain/langgraph @langchain/anthropic @langchain/core

初始化 LLM

import { ChatAnthropic } from "@langchain/anthropic";

process.env.ANTHROPIC_API_KEY = "<your_anthropic_key>";

const llm = new ChatAnthropic({
  model: "claude-3-5-sonnet-latest",
});

构建模块：增强型 LLM¶

LLM 具有支持构建工作流和 Agent 的增强功能。这些功能包括结构化输出和工具调用，如下图所示，来自 Anthropic 博客

import { tool } from "@langchain/core/tools";
import { z } from "zod";

const searchQuerySchema = z.object({
  searchQuery: z.string().describe("Query that is optimized web search."),
  justification: z.string("Why this query is relevant to the user's request."),
});

// Augment the LLM with schema for structured output
const structuredLlm = llm.withStructuredOutput(searchQuerySchema, {
  name: "searchQuery",
});

// Invoke the augmented LLM
const output = await structuredLlm.invoke(
  "How does Calcium CT score relate to high cholesterol?"
);

const multiply = tool(
  async ({ a, b }) => {
    return a * b;
  },
  {
    name: "multiply",
    description: "multiplies two numbers together",
    schema: z.object({
      a: z.number("the first number"),
      b: z.number("the second number"),
    }),
  }
);

// Augment the LLM with tools
const llmWithTools = llm.bindTools([multiply]);

// Invoke the LLM with input that triggers the tool call
const message = await llmWithTools.invoke("What is 2 times 3?");

console.log(message.tool_calls);

Prompt 链¶

在 Prompt 链中，每个 LLM 调用处理前一个调用的输出。

正如 Anthropic 博客中指出的那样

Prompt 链将任务分解为一系列步骤，其中每个 LLM 调用处理前一个调用的输出。您可以在任何中间步骤添加程序化检查（请参阅下图中的“gate”），以确保流程仍在轨道上。

何时使用此工作流：此工作流非常适合任务可以轻松干净地分解为固定子任务的情况。主要目标是通过使每个 LLM 调用成为更简单的任务，来权衡延迟以获得更高的准确性。

Graph APIFunctional API (beta)

import { StateGraph, Annotation } from "@langchain/langgraph";

// Graph state
const StateAnnotation = Annotation.Root({
  topic: Annotation<string>,
  joke: Annotation<string>,
  improvedJoke: Annotation<string>,
  finalJoke: Annotation<string>,
});

// Define node functions

// First LLM call to generate initial joke
async function generateJoke(state: typeof StateAnnotation.State) {
  const msg = await llm.invoke(`Write a short joke about ${state.topic}`);
  return { joke: msg.content };
}

// Gate function to check if the joke has a punchline
function checkPunchline(state: typeof StateAnnotation.State) {
  // Simple check - does the joke contain "?" or "!"
  if (state.joke?.includes("?") || state.joke?.includes("!")) {
    return "Pass";
  }
  return "Fail";
}

  // Second LLM call to improve the joke
async function improveJoke(state: typeof StateAnnotation.State) {
  const msg = await llm.invoke(
    `Make this joke funnier by adding wordplay: ${state.joke}`
  );
  return { improvedJoke: msg.content };
}

// Third LLM call for final polish
async function polishJoke(state: typeof StateAnnotation.State) {
  const msg = await llm.invoke(
    `Add a surprising twist to this joke: ${state.improvedJoke}`
  );
  return { finalJoke: msg.content };
}

// Build workflow
const chain = new StateGraph(StateAnnotation)
  .addNode("generateJoke", generateJoke)
  .addNode("improveJoke", improveJoke)
  .addNode("polishJoke", polishJoke)
  .addEdge("__start__", "generateJoke")
  .addConditionalEdges("generateJoke", checkPunchline, {
    Pass: "improveJoke",
    Fail: "__end__"
  })
  .addEdge("improveJoke", "polishJoke")
  .addEdge("polishJoke", "__end__")
  .compile();

// Invoke
const state = await chain.invoke({ topic: "cats" });
console.log("Initial joke:");
console.log(state.joke);
console.log("\n--- --- ---\n");
if (state.improvedJoke !== undefined) {
  console.log("Improved joke:");
  console.log(state.improvedJoke);
  console.log("\n--- --- ---\n");

  console.log("Final joke:");
  console.log(state.finalJoke);
} else {
  console.log("Joke failed quality gate - no punchline detected!");
}

LangSmith Trace

https://smith.langchain.com/public/a0281fca-3a71-46de-beee-791468607b75/r

import { task, entrypoint } from "@langchain/langgraph";

// Tasks

// First LLM call to generate initial joke
const generateJoke = task("generateJoke", async (topic: string) => {
  const msg = await llm.invoke(`Write a short joke about ${topic}`);
  return msg.content;
});

// Gate function to check if the joke has a punchline
function checkPunchline(joke: string) {
  // Simple check - does the joke contain "?" or "!"
  if (joke.includes("?") || joke.includes("!")) {
    return "Pass";
  }
  return "Fail";
}

  // Second LLM call to improve the joke
const improveJoke = task("improveJoke", async (joke: string) => {
  const msg = await llm.invoke(
    `Make this joke funnier by adding wordplay: ${joke}`
  );
  return msg.content;
});

// Third LLM call for final polish
const polishJoke = task("polishJoke", async (joke: string) => {
  const msg = await llm.invoke(
    `Add a surprising twist to this joke: ${joke}`
  );
  return msg.content;
});

const workflow = entrypoint(
  "jokeMaker",
  async (topic: string) => {
    const originalJoke = await generateJoke(topic);
    if (checkPunchline(originalJoke) === "Pass") {
      return originalJoke;
    }
    const improvedJoke = await improveJoke(originalJoke);
    const polishedJoke = await polishJoke(improvedJoke);
    return polishedJoke;
  }
);

const stream = await workflow.stream("cats", {
  streamMode: "updates",
});

for await (const step of stream) {
  console.log(step);
}

LangSmith Trace

https://smith.langchain.com/public/332fa4fc-b6ca-416e-baa3-161625e69163/r

并行化¶

通过并行化，LLM 同时处理一项任务

LLM 有时可以同时处理一项任务，并以编程方式聚合其输出。这种工作流，并行化，主要体现在两种关键变体中：分段：将任务分解为并行运行的独立子任务。投票：多次运行同一任务以获得不同的输出。

何时使用此工作流：当可以并行化划分的子任务以提高速度，或者当需要多个视角或尝试以获得更高置信度的结果时，并行化是有效的。对于具有多个考虑因素的复杂任务，通常当每个考虑因素由单独的 LLM 调用处理时，LLM 的性能会更好，从而可以集中关注每个特定方面。

Graph APIFunctional API (beta)

import { StateGraph, Annotation } from "@langchain/langgraph";

// Graph state
const StateAnnotation = Annotation.Root({
  topic: Annotation<string>,
  joke: Annotation<string>,
  story: Annotation<string>,
  poem: Annotation<string>,
  combinedOutput: Annotation<string>,
});

// Nodes
// First LLM call to generate initial joke
async function callLlm1(state: typeof StateAnnotation.State) {
  const msg = await llm.invoke(`Write a joke about ${state.topic}`);
  return { joke: msg.content };
}

// Second LLM call to generate story
async function callLlm2(state: typeof StateAnnotation.State) {
  const msg = await llm.invoke(`Write a story about ${state.topic}`);
  return { story: msg.content };
}

// Third LLM call to generate poem
async function callLlm3(state: typeof StateAnnotation.State) {
  const msg = await llm.invoke(`Write a poem about ${state.topic}`);
  return { poem: msg.content };
}

// Combine the joke, story and poem into a single output
async function aggregator(state: typeof StateAnnotation.State) {
  const combined = `Here's a story, joke, and poem about ${state.topic}!\n\n` +
    `STORY:\n${state.story}\n\n` +
    `JOKE:\n${state.joke}\n\n` +
    `POEM:\n${state.poem}`;
  return { combinedOutput: combined };
}

// Build workflow
const parallelWorkflow = new StateGraph(StateAnnotation)
  .addNode("callLlm1", callLlm1);
  .addNode("callLlm2", callLlm2);
  .addNode("callLlm3", callLlm3);
  .addNode("aggregator", aggregator);
  .addEdge("__start__", "callLlm1");
  .addEdge("__start__", "callLlm2");
  .addEdge("__start__", "callLlm3");
  .addEdge("callLlm1", "aggregator");
  .addEdge("callLlm2", "aggregator");
  .addEdge("callLlm3", "aggregator");
  .addEdge("aggregator", "__end__");
  .compile();

// Invoke
const result = await parallelWorkflow.invoke({ topic: "cats" });
console.log(result.combinedOutput);

LangSmith Trace

https://smith.langchain.com/public/3be2e53c-ca94-40dd-934f-82ff87fac277/r

资源

文档

请参阅我们关于并行化的文档这里。

import { task, entrypoint } from "@langchain/langgraph";

// Tasks

// First LLM call to generate initial joke
const callLlm1 = task("generateJoke", async (topic: string) => {
  const msg = await llm.invoke(`Write a joke about ${topic}`);
  return msg.content;
});

// Second LLM call to generate story
const callLlm2 = task("generateStory", async (topic: string) => {
  const msg = await llm.invoke(`Write a story about ${topic}`);
  return msg.content;
});

// Third LLM call to generate poem
const callLlm3 = task("generatePoem", async (topic: string) => {
  const msg = await llm.invoke(`Write a poem about ${topic}`);
  return msg.content;
});

// Combine outputs
const aggregator = task("aggregator", async (params: {
  topic: string;
  joke: string;
  story: string;
  poem: string;
}) => {
  const { topic, joke, story, poem } = params;
  return `Here's a story, joke, and poem about ${topic}!\n\n` +
    `STORY:\n${story}\n\n` +
    `JOKE:\n${joke}\n\n` +
    `POEM:\n${poem}`;
});

// Build workflow
const workflow = entrypoint(
  "parallelWorkflow",
  async (topic: string) => {
    const [joke, story, poem] = await Promise.all([
      callLlm1(topic),
      callLlm2(topic),
      callLlm3(topic),
    ]);

    return aggregator({ topic, joke, story, poem });
  }
);

// Invoke
const stream = await workflow.stream("cats", {
  streamMode: "updates",
});

for await (const step of stream) {
  console.log(step);
}

LangSmith Trace

https://smith.langchain.com/public/623d033f-e814-41e9-80b1-75e6abb67801/r

路由¶

路由对输入进行分类，并将其定向到后续任务。正如 Anthropic 博客中指出的那样

路由对输入进行分类，并将其定向到专门的后续任务。这种工作流允许关注点分离，并构建更专业的 Prompt。如果没有这种工作流，优化一种输入可能会损害其他输入的性能。

何时使用此工作流：路由非常适用于复杂任务，在这些任务中，存在最好单独处理的不同类别，并且可以通过 LLM 或更传统的分类模型/算法准确处理分类。

Graph APIFunctional API (beta)

import { StateGraph, Annotation } from "@langchain/langgraph";
import { z } from "zod";

// Schema for structured output to use as routing logic
const routeSchema = z.object({
  step: z.enum(["poem", "story", "joke"]).describe(
    "The next step in the routing process"
  ),
});

// Augment the LLM with schema for structured output
const router = llm.withStructuredOutput(routeSchema);

// Graph state
const StateAnnotation = Annotation.Root({
  input: Annotation<string>,
  decision: Annotation<string>,
  output: Annotation<string>,
});

// Nodes
// Write a story
async function llmCall1(state: typeof StateAnnotation.State) {
  const result = await llm.invoke([{
    role: "system",
    content: "You are an expert storyteller.",
  }, {
    role: "user",
    content: state.input
  }]);
  return { output: result.content };
}

// Write a joke
async function llmCall2(state: typeof StateAnnotation.State) {
  const result = await llm.invoke([{
    role: "system",
    content: "You are an expert comedian.",
  }, {
    role: "user",
    content: state.input
  }]);
  return { output: result.content };
}

// Write a poem
async function llmCall3(state: typeof StateAnnotation.State) {
  const result = await llm.invoke([{
    role: "system",
    content: "You are an expert poet.",
  }, {
    role: "user",
    content: state.input
  }]);
  return { output: result.content };
}

async function llmCallRouter(state: typeof StateAnnotation.State) {
  // Route the input to the appropriate node
  const decision = await router.invoke([
    {
      role: "system",
      content: "Route the input to story, joke, or poem based on the user's request."
    },
    {
      role: "user",
      content: state.input
    },
  ]);

  return { decision: decision.step };
}

// Conditional edge function to route to the appropriate node
function routeDecision(state: typeof StateAnnotation.State) {
  // Return the node name you want to visit next
  if (state.decision === "story") {
    return "llmCall1";
  } else if (state.decision === "joke") {
    return "llmCall2";
  } else if (state.decision === "poem") {
    return "llmCall3";
  }
}

// Build workflow
const routerWorkflow = new StateGraph(StateAnnotation)
  .addNode("llmCall1", llmCall1)
  .addNode("llmCall2", llmCall2)
  .addNode("llmCall3", llmCall3)
  .addNode("llmCallRouter", llmCallRouter)
  .addEdge("__start__", "llmCallRouter")
  .addConditionalEdges(
    "llmCallRouter",
    routeDecision,
    ["llmCall1", "llmCall2", "llmCall3"],
  )
  .addEdge("llmCall1", "__end__")
  .addEdge("llmCall2", "__end__")
  .addEdge("llmCall3", "__end__")
  .compile();

// Invoke
const state = await routerWorkflow.invoke({
  input: "Write me a joke about cats"
});
console.log(state.output);

LangSmith Trace

https://smith.langchain.com/public/c4580b74-fe91-47e4-96fe-7fac598d509c/r

示例

这里是一个 RAG 工作流，用于路由问题。请观看我们的视频这里。

import { z } from "zod";
import { task, entrypoint } from "@langchain/langgraph";

// Schema for structured output to use as routing logic
const routeSchema = z.object({
  step: z.enum(["poem", "story", "joke"]).describe(
    "The next step in the routing process"
  ),
});

// Augment the LLM with schema for structured output
const router = llm.withStructuredOutput(routeSchema);

// Tasks
// Write a story
const llmCall1 = task("generateStory", async (input: string) => {
  const result = await llm.invoke([{
    role: "system",
    content: "You are an expert storyteller.",
  }, {
    role: "user",
    content: input
  }]);
  return result.content;
});

// Write a joke
const llmCall2 = task("generateJoke", async (input: string) => {
  const result = await llm.invoke([{
    role: "system",
    content: "You are an expert comedian.",
  }, {
    role: "user",
    content: input
  }]);
  return result.content;
});

// Write a poem
const llmCall3 = task("generatePoem", async (input: string) => {
  const result = await llm.invoke([{
    role: "system",
    content: "You are an expert poet.",
  }, {
    role: "user",
    content: input
  }]);
  return result.content;
});

// Route the input to the appropriate node
const llmCallRouter = task("router", async (input: string) => {
  const decision = await router.invoke([
    {
      role: "system",
      content: "Route the input to story, joke, or poem based on the user's request."
    },
    {
      role: "user",
      content: input
    },
  ]);
  return decision.step;
});

// Build workflow
const workflow = entrypoint(
  "routerWorkflow",
  async (input: string) => {
    const nextStep = await llmCallRouter(input);

    let llmCall;
    if (nextStep === "story") {
      llmCall = llmCall1;
    } else if (nextStep === "joke") {
      llmCall = llmCall2;
    } else if (nextStep === "poem") {
      llmCall = llmCall3;
    }

    const finalResult = await llmCall(input);
    return finalResult;
  }
);

// Invoke
const stream = await workflow.stream("Write me a joke about cats", {
  streamMode: "updates",
});

for await (const step of stream) {
  console.log(step);
}

LangSmith Trace

https://smith.langchain.com/public/5e2eb979-82dd-402c-b1a0-a8cceaf2a28a/r

编排器-工作器¶

通过编排器-工作器，编排器分解任务并将每个子任务委派给工作器。正如 Anthropic 博客中指出的那样

在编排器-工作器工作流中，中央 LLM 动态分解任务，将其委派给工作器 LLM，并综合它们的结果。

何时使用此工作流：此工作流非常适合您无法预测所需子任务的复杂任务（例如，在编码中，需要更改的文件数量以及每个文件中更改的性质可能取决于任务）。虽然它在地形上相似，但与并行化的关键区别在于其灵活性——子任务不是预定义的，而是由编排器根据特定输入确定的。

Graph APIFunctional API (beta)

import { z } from "zod";

// Schema for structured output to use in planning
const sectionSchema = z.object({
  name: z.string().describe("Name for this section of the report."),
  description: z.string().describe(
    "Brief overview of the main topics and concepts to be covered in this section."
  ),
});

const sectionsSchema = z.object({
  sections: z.array(sectionSchema).describe("Sections of the report."),
});

// Augment the LLM with schema for structured output
const planner = llm.withStructuredOutput(sectionsSchema);

在 LangGraph 中创建工作器

由于编排器-工作器工作流很常见，LangGraph 具有 Send API 来支持这一点。它允许您动态创建工作器节点，并向每个节点发送特定的输入。每个工作器都有自己的状态，所有工作器输出都写入一个共享状态键，编排器图可以访问该键。这使编排器可以访问所有工作器输出，并允许它将它们合成为最终输出。正如您在下面看到的，我们迭代一个部分列表，并将每个部分 Send 到一个工作器节点。请参阅更多文档这里和这里。

import { Annotation, StateGraph, Send } from "@langchain/langgraph";

// Graph state
const StateAnnotation = Annotation.Root({
  topic: Annotation<string>,
  sections: Annotation<Array<z.infer<typeof sectionSchema>>>,
  completedSections: Annotation<string[]>({
    default: () => [],
    reducer: (a, b) => a.concat(b),
  }),
  finalReport: Annotation<string>,
});

// Worker state
const WorkerStateAnnotation = Annotation.Root({
  section: Annotation<z.infer<typeof sectionSchema>>,
  completedSections: Annotation<string[]>({
    default: () => [],
    reducer: (a, b) => a.concat(b),
  }),
});

// Nodes
async function orchestrator(state: typeof StateAnnotation.State) {
  // Generate queries
  const reportSections = await planner.invoke([
    { role: "system", content: "Generate a plan for the report." },
    { role: "user", content: `Here is the report topic: ${state.topic}` },
  ]);

  return { sections: reportSections.sections };
}

async function llmCall(state: typeof WorkerStateAnnotation.State) {
  // Generate section
  const section = await llm.invoke([
    {
      role: "system",
      content: "Write a report section following the provided name and description. Include no preamble for each section. Use markdown formatting.",
    },
    {
      role: "user",
      content: `Here is the section name: ${state.section.name} and description: ${state.section.description}`,
    },
  ]);

  // Write the updated section to completed sections
  return { completedSections: [section.content] };
}

async function synthesizer(state: typeof StateAnnotation.State) {
  // List of completed sections
  const completedSections = state.completedSections;

  // Format completed section to str to use as context for final sections
  const completedReportSections = completedSections.join("\n\n---\n\n");

  return { finalReport: completedReportSections };
}

// Conditional edge function to create llm_call workers that each write a section of the report
function assignWorkers(state: typeof StateAnnotation.State) {
  // Kick off section writing in parallel via Send() API
  return state.sections.map((section) =>
    new Send("llmCall", { section })
  );
}

// Build workflow
const orchestratorWorker = new StateGraph(StateAnnotation)
  .addNode("orchestrator", orchestrator)
  .addNode("llmCall", llmCall)
  .addNode("synthesizer", synthesizer)
  .addEdge("__start__", "orchestrator")
  .addConditionalEdges(
    "orchestrator",
    assignWorkers,
    ["llmCall"]
  )
  .addEdge("llmCall", "synthesizer")
  .addEdge("synthesizer", "__end__")
  .compile();

// Invoke
const state = await orchestratorWorker.invoke({
  topic: "Create a report on LLM scaling laws"
});
console.log(state.finalReport);

LangSmith Trace

https://smith.langchain.com/public/78cbcfc3-38bf-471d-b62a-b299b144237d/r

资源

示例

这里是一个使用编排器-工作器进行报告计划和编写的项目。请观看我们的视频这里。

import { z } from "zod";
import { task, entrypoint } from "@langchain/langgraph";

// Schema for structured output to use in planning
const sectionSchema = z.object({
  name: z.string().describe("Name for this section of the report."),
  description: z.string().describe(
    "Brief overview of the main topics and concepts to be covered in this section."
  ),
});

const sectionsSchema = z.object({
  sections: z.array(sectionSchema).describe("Sections of the report."),
});

// Augment the LLM with schema for structured output
const planner = llm.withStructuredOutput(sectionsSchema);

// Tasks
const orchestrator = task("orchestrator", async (topic: string) => {
  // Generate queries
  const reportSections = await planner.invoke([
    { role: "system", content: "Generate a plan for the report." },
    { role: "user", content: `Here is the report topic: ${topic}` },
  ]);

  return reportSections.sections;
});

const llmCall = task("sectionWriter", async (section: z.infer<typeof sectionSchema>) => {
  // Generate section
  const result = await llm.invoke([
    {
      role: "system",
      content: "Write a report section.",
    },
    {
      role: "user",
      content: `Here is the section name: ${section.name} and description: ${section.description}`,
    },
  ]);

  return result.content;
});

const synthesizer = task("synthesizer", async (completedSections: string[]) => {
  // Synthesize full report from sections
  return completedSections.join("\n\n---\n\n");
});

// Build workflow
const workflow = entrypoint(
  "orchestratorWorker",
  async (topic: string) => {
    const sections = await orchestrator(topic);
    const completedSections = await Promise.all(
      sections.map((section) => llmCall(section))
    );
    return synthesizer(completedSections);
  }
);

// Invoke
const stream = await workflow.stream("Create a report on LLM scaling laws", {
  streamMode: "updates",
});

for await (const step of stream) {
  console.log(step);
}

LangSmith Trace

https://smith.langchain.com/public/75a636d0-6179-4a12-9836-e0aa571e87c5/r

评估器-优化器¶

在评估器-优化器工作流中，一个 LLM 调用生成响应，而另一个 LLM 调用在循环中提供评估和反馈

在评估器-优化器工作流中，一个 LLM 调用生成响应，而另一个 LLM 调用在循环中提供评估和反馈。

何时使用此工作流：当我们有明确的评估标准，并且迭代改进提供可衡量的价值时，此工作流尤其有效。良好契合的两个迹象是，首先，当人表达他们的反馈时，LLM 响应可以得到明显的改进；其次，LLM 可以提供此类反馈。这类似于人类作者在制作润色文档时可能经历的迭代写作过程。

Graph APIFunctional API (beta)

import { z } from "zod";
import { Annotation, StateGraph } from "@langchain/langgraph";

// Graph state
const StateAnnotation = Annotation.Root({
  joke: Annotation<string>,
  topic: Annotation<string>,
  feedback: Annotation<string>,
  funnyOrNot: Annotation<string>,
});

// Schema for structured output to use in evaluation
const feedbackSchema = z.object({
  grade: z.enum(["funny", "not funny"]).describe(
    "Decide if the joke is funny or not."
  ),
  feedback: z.string().describe(
    "If the joke is not funny, provide feedback on how to improve it."
  ),
});

// Augment the LLM with schema for structured output
const evaluator = llm.withStructuredOutput(feedbackSchema);

// Nodes
async function llmCallGenerator(state: typeof StateAnnotation.State) {
  // LLM generates a joke
  let msg;
  if (state.feedback) {
    msg = await llm.invoke(
      `Write a joke about ${state.topic} but take into account the feedback: ${state.feedback}`
    );
  } else {
    msg = await llm.invoke(`Write a joke about ${state.topic}`);
  }
  return { joke: msg.content };
}

async function llmCallEvaluator(state: typeof StateAnnotation.State) {
  // LLM evaluates the joke
  const grade = await evaluator.invoke(`Grade the joke ${state.joke}`);
  return { funnyOrNot: grade.grade, feedback: grade.feedback };
}

// Conditional edge function to route back to joke generator or end based upon feedback from the evaluator
function routeJoke(state: typeof StateAnnotation.State) {
  // Route back to joke generator or end based upon feedback from the evaluator
  if (state.funnyOrNot === "funny") {
    return "Accepted";
  } else if (state.funnyOrNot === "not funny") {
    return "Rejected + Feedback";
  }
}

// Build workflow
const optimizerWorkflow = new StateGraph(StateAnnotation)
  .addNode("llmCallGenerator", llmCallGenerator)
  .addNode("llmCallEvaluator", llmCallEvaluator)
  .addEdge("__start__", "llmCallGenerator")
  .addEdge("llmCallGenerator", "llmCallEvaluator")
  .addConditionalEdges(
    "llmCallEvaluator",
    routeJoke,
    {
      // Name returned by routeJoke : Name of next node to visit
      "Accepted": "__end__",
      "Rejected + Feedback": "llmCallGenerator",
    }
  )
  .compile();

// Invoke
const state = await optimizerWorkflow.invoke({ topic: "Cats" });
console.log(state.joke);

LangSmith Trace

https://smith.langchain.com/public/86ab3e60-2000-4bff-b988-9b89a3269789/r

资源

示例

这里是一个使用评估器-优化器来改进报告的助手。请观看我们的视频这里。

这里是一个 RAG 工作流，用于评估答案是否存在幻觉或错误。请观看我们的视频这里。

import { z } from "zod";
import { task, entrypoint } from "@langchain/langgraph";

// Schema for structured output to use in evaluation
const feedbackSchema = z.object({
  grade: z.enum(["funny", "not funny"]).describe(
    "Decide if the joke is funny or not."
  ),
  feedback: z.string().describe(
    "If the joke is not funny, provide feedback on how to improve it."
  ),
});

// Augment the LLM with schema for structured output
const evaluator = llm.withStructuredOutput(feedbackSchema);

// Tasks
const llmCallGenerator = task("jokeGenerator", async (params: {
  topic: string;
  feedback?: z.infer<typeof feedbackSchema>;
}) => {
  // LLM generates a joke
  const msg = params.feedback
    ? await llm.invoke(
        `Write a joke about ${params.topic} but take into account the feedback: ${params.feedback.feedback}`
      )
    : await llm.invoke(`Write a joke about ${params.topic}`);
  return msg.content;
});

const llmCallEvaluator = task("jokeEvaluator", async (joke: string) => {
  // LLM evaluates the joke
  return evaluator.invoke(`Grade the joke ${joke}`);
});

// Build workflow
const workflow = entrypoint(
  "optimizerWorkflow",
  async (topic: string) => {
    let feedback: z.infer<typeof feedbackSchema> | undefined;
    let joke: string;

    while (true) {
      joke = await llmCallGenerator({ topic, feedback });
      feedback = await llmCallEvaluator(joke);

      if (feedback.grade === "funny") {
        break;
      }
    }

    return joke;
  }
);

// Invoke
const stream = await workflow.stream("Cats", {
  streamMode: "updates",
});

for await (const step of stream) {
  console.log(step);
  console.log("\n");
}

LangSmith Trace

https://smith.langchain.com/public/f66830be-4339-4a6b-8a93-389ce5ae27b4/r

Agent¶

Agent 通常实现为 LLM 基于循环中的环境反馈执行操作（通过工具调用）。正如 Anthropic 博客中指出的那样

Agent 可以处理复杂的任务，但它们的实现通常很简单。它们通常只是 LLM 基于循环中的环境反馈使用工具。因此，至关重要的是要清晰且周到地设计工具集及其文档。

何时使用 Agent：Agent 可用于开放式问题，在这些问题中，很难或不可能预测所需的步骤数，并且您无法硬编码固定路径。LLM 可能会运行很多轮，您必须对其决策制定有一定的信任度。Agent 的自主性使其成为在受信任环境中扩展任务的理想选择。

import { tool } from "@langchain/core/tools";
import { z } from "zod";

// Define tools
const multiply = tool(
  async ({ a, b }: { a: number; b: number }) => {
    return a * b;
  },
  {
    name: "multiply",
    description: "Multiply two numbers together",
    schema: z.object({
      a: z.number().describe("first number"),
      b: z.number().describe("second number"),
    }),
  }
);

const add = tool(
  async ({ a, b }: { a: number; b: number }) => {
    return a + b;
  },
  {
    name: "add",
    description: "Add two numbers together",
    schema: z.object({
      a: z.number().describe("first number"),
      b: z.number().describe("second number"),
    }),
  }
);

const divide = tool(
  async ({ a, b }: { a: number; b: number }) => {
    return a / b;
  },
  {
    name: "divide",
    description: "Divide two numbers",
    schema: z.object({
      a: z.number().describe("first number"),
      b: z.number().describe("second number"),
    }),
  }
);

// Augment the LLM with tools
const tools = [add, multiply, divide];
const toolsByName = Object.fromEntries(tools.map((tool) => [tool.name, tool]));
const llmWithTools = llm.bindTools(tools);

Graph APIFunctional API (beta)

import { MessagesAnnotation, StateGraph } from "@langchain/langgraph";
import { ToolNode } from "@langchain/langgraph/prebuilt";
import {
  SystemMessage,
  ToolMessage
} from "@langchain/core/messages";

// Nodes
async function llmCall(state: typeof MessagesAnnotation.State) {
  // LLM decides whether to call a tool or not
  const result = await llmWithTools.invoke([
    {
      role: "system",
      content: "You are a helpful assistant tasked with performing arithmetic on a set of inputs."
    },
    ...state.messages
  ]);

  return {
    messages: [result]
  };
}

const toolNode = new ToolNode(tools);

// Conditional edge function to route to the tool node or end
function shouldContinue(state: typeof MessagesAnnotation.State) {
  const messages = state.messages;
  const lastMessage = messages.at(-1);

  // If the LLM makes a tool call, then perform an action
  if (lastMessage?.tool_calls?.length) {
    return "Action";
  }
  // Otherwise, we stop (reply to the user)
  return "__end__";
}

// Build workflow
const agentBuilder = new StateGraph(MessagesAnnotation)
  .addNode("llmCall", llmCall)
  .addNode("tools", toolNode)
  // Add edges to connect nodes
  .addEdge("__start__", "llmCall")
  .addConditionalEdges(
    "llmCall",
    shouldContinue,
    {
      // Name returned by shouldContinue : Name of next node to visit
      "Action": "tools",
      "__end__": "__end__",
    }
  )
  .addEdge("tools", "llmCall")
  .compile();

// Invoke
const messages = [{
  role: "user",
  content: "Add 3 and 4."
}];
const result = await agentBuilder.invoke({ messages });
console.log(result.messages);

LangSmith Trace

https://smith.langchain.com/public/051f0391-6761-4f8c-a53b-22231b016690/r

示例

这里是一个使用工具调用 Agent 来创建/存储长期记忆的项目。

import { task, entrypoint, addMessages } from "@langchain/langgraph";
import { BaseMessageLike, ToolCall } from "@langchain/core/messages";

const callLlm = task("llmCall", async (messages: BaseMessageLike[]) => {
  // LLM decides whether to call a tool or not
  return llmWithTools.invoke([
    {
      role: "system",
      content: "You are a helpful assistant tasked with performing arithmetic on a set of inputs."
    },
    ...messages
  ]);
});

const callTool = task("toolCall", async (toolCall: ToolCall) => {
  // Performs the tool call
  const tool = toolsByName[toolCall.name];
  return tool.invoke(toolCall.args);
});

const agent = entrypoint(
  "agent",
  async (messages: BaseMessageLike[]) => {
    let llmResponse = await callLlm(messages);

    while (true) {
      if (!llmResponse.tool_calls?.length) {
        break;
      }

      // Execute tools
      const toolResults = await Promise.all(
        llmResponse.tool_calls.map((toolCall) => callTool(toolCall))
      );

      messages = addMessages(messages, [llmResponse, ...toolResults]);
      llmResponse = await callLlm(messages);
    }

    messages = addMessages(messages, [llmResponse]);
    return messages;
  }
);

// Invoke
const messages = [{
  role: "user",
  content: "Add 3 and 4."
}];

const stream = await agent.stream([messages], {
  streamMode: "updates",
});

for await (const step of stream) {
  console.log(step);
}

LangSmith Trace

https://smith.langchain.com/public/42ae8bf9-3935-4504-a081-8ddbcbfc8b2e/r

预构建¶

LangGraph 还提供了一种预构建方法，用于创建如上定义的 Agent（使用 createReactAgent 函数）

https://github.langchain.ac.cn/langgraphjs/how-tos/create-react-agent/

import { createReactAgent } from "@langchain/langgraph/prebuilt";

// Pass in:
// (1) an LLM instance
// (2) the tools list (which is used to create the tool node)
const prebuiltAgent = createReactAgent({
  llm: llmWithTools,
  tools,
});

// invoke
const result = await prebuiltAgent.invoke({
  messages: [
    {
      role: "user",
      content: "Add 3 and 4.",
    },
  ],
});
console.log(result.messages);

LangSmith Trace

https://smith.langchain.com/public/abab6a44-29f6-4b97-8164-af77413e494d/r

LangGraph 提供的功能¶

通过在 LangGraph 中构建上述每项功能，我们获得了一些东西

持久性：人机环路¶

LangGraph 持久层支持操作的中断和批准（例如，人机环路）。请参阅 LangChain 学院的模块 3。

持久性：记忆¶

LangGraph 持久层支持对话式（短期）记忆和长期记忆。请参阅 LangChain 学院的模块 2 和 5

流式传输¶

LangGraph 提供了几种流式传输工作流/Agent 输出或中间状态的方法。请参阅 LangChain 学院的模块 3。

部署¶

LangGraph 为部署、可观察性和评估提供了简单的入门途径。请参阅 LangChain 学院的模块 6。