尚拙
Search K

Appearance

结构化输出 (Structured output) ​

结构化输出允许智能体以特定、可预测的格式返回数据。无需解析自然语言响应，你即可获得带类型的结构化数据。

TIP

本页介绍在结合 createAgent 的智能体中使用结构化输出。如果要在模型上直接使用结构化输出（在智能体之外），请参阅 模型 - 结构化输出。

LangChain 的预置 ReAct 智能体 createAgent 会自动处理结构化输出。用户设置所需的结构化输出 Schema，当模型生成结构化数据时，该数据会被捕获、校验，并在智能体状态的 structuredResponse 键中返回。

ts
type ResponseFormat = (
    | ZodSchema<StructuredResponseT> // Zod schema
    | StandardSchema<StructuredResponseT> // 任何 Standard Schema 兼容库
    | Record<string, unknown> // JSON Schema
)

const agent = createAgent({
    // ...
    responseFormat: ResponseFormat | ResponseFormat[]
})

响应格式 (Response format) ​

控制智能体如何返回结构化数据。你可以提供 Zod schema、任何兼容 Standard Schema 的 schema，或 JSON Schema 对象。默认情况下，智能体使用工具调用策略，在此策略下，输出是通过一次额外的工具调用创建的。某些模型支持原生结构化输出，在这种情况下智能体将改用该策略。

你可以通过将 ResponseFormat 包装在 toolStrategy 或 providerStrategy 函数调用中来控制该行为：

ts
import { toolStrategy, providerStrategy } from "langchain";

const agent = createAgent({
    // 如果模型支持，则使用提供商策略
    responseFormat: providerStrategy(z.object({ ... }))
    // 或者强制使用工具策略
    responseFormat: toolStrategy(z.object({ ... }))
})

结构化响应将在智能体最终状态的 structuredResponse 键中返回。

TIP

如果使用的是 langchain>=1.1，对原生结构化输出功能的支持会从模型的 Profile 数据 中动态读取。如果数据不可用，可以使用其他条件或手动指定：

typescript
const customProfile: ModelProfile = {
    structuredOutput: true,
    // ...
}
const model = await initChatModel("...", { profile: customProfile });

如果指定了工具，模型必须支持同时使用工具和结构化输出。

提供商策略 (Provider strategy) ​

某些模型提供商通过其 API 原生支持结构化输出（例如 OpenAI、xAI (Grok)、Gemini、Anthropic (Claude)）。在可用时，这是最可靠的方法。

要使用此策略，请配置 ProviderStrategy：

ts
function providerStrategy<StructuredResponseT>(
    schema: ZodSchema<StructuredResponseT> | SerializableSchema | JsonSchemaFormat
): ProviderStrategy<StructuredResponseT>

schema 必填

定义结构化输出格式的 Schema。支持：

  • Zod Schema：Zod schema
  • Standard Schema：任何实现 Standard Schema 规范的 schema
  • JSON Schema：JSON Schema 对象

当你直接将 schema 类型传递给 createAgent.responseFormat 且模型支持原生结构化输出时，LangChain 会自动使用 ProviderStrategy：

ts
import * as z from "zod";
import { createAgent, providerStrategy } from "langchain";

const ContactInfo = z.object({
    name: z.string().describe("The name of the person"),
    email: z.string().describe("The email address of the person"),
    phone: z.string().describe("The phone number of the person"),
});

const agent = createAgent({
    model: "gpt-5.5",
    tools: [],
    responseFormat: providerStrategy(ContactInfo)
});

const result = await agent.invoke({
    messages: [{"role": "user", "content": "Extract contact info from: John Doe, john@example.com, (555) 123-4567"}]
});

console.log(result.structuredResponse);
// { name: "John Doe", email: "john@example.com", phone: "(555) 123-4567" }
ts
import * as v from "valibot";
import { toStandardJsonSchema } from "@valibot/to-json-schema";
import { createAgent, providerStrategy } from "langchain";

const ContactInfo = toStandardJsonSchema(
    v.object({
        name: v.pipe(v.string(), v.description("The name of the person")),
        email: v.pipe(v.string(), v.description("The email address of the person")),
        phone: v.pipe(v.string(), v.description("The phone number of the person")),
    })
);

const agent = createAgent({
    model: "gpt-5.5",
    tools: [],
    responseFormat: providerStrategy(ContactInfo)
});

const result = await agent.invoke({
    messages: [{"role": "user", "content": "Extract contact info from: John Doe, john@example.com, (555) 123-4567"}]
});

console.log(result.structuredResponse);
// { name: "John Doe", email: "john@example.com", phone: "(555) 123-4567" }
ts
import { createAgent, providerStrategy } from "langchain";

const contactInfoSchema = {
    "type": "object",
    "description": "Contact information for a person.",
    "properties": {
        "name": {"type": "string", "description": "The name of the person"},
        "email": {"type": "string", "description": "The email address of the person"},
        "phone": {"type": "string", "description": "The phone number of the person"}
    },
    "required": ["name", "email", "phone"]
}

const agent = createAgent({
    model: "gpt-5.5",
    tools: [],
    responseFormat: providerStrategy(contactInfoSchema)
});

const result = await agent.invoke({
    messages: [{"role": "user", "content": "Extract contact info from: John Doe, john@example.com, (555) 123-4567"}]
});

console.log(result.structuredResponse);
// { name: "John Doe", email: "john@example.com", phone: "(555) 123-4567" }

提供商原生结构化输出提供了极高的可靠性和严格的校验，因为模型提供商会强制执行该 Schema。在可用时建议优先使用。

NOTE

如果你所选择的模型提供商原生支持结构化输出，直接编写 responseFormat: contactInfoSchema 与编写 responseFormat: providerStrategy(contactInfoSchema) 在功能上是等效的。

在任何一种情况下，如果不支持结构化输出，智能体都将回退到工具调用策略。

工具调用策略 (Tool calling strategy) ​

对于不支持原生结构化输出的模型，LangChain 使用工具调用来实现相同的效果。这适用于所有支持工具调用的模型（大多数现代模型）。

要使用此策略，请配置 ToolStrategy：

ts
function toolStrategy<StructuredResponseT>(
    responseFormat:
        | JsonSchemaFormat
        | ZodSchema<StructuredResponseT>
        | SerializableSchema
        | (ZodSchema<StructuredResponseT> | SerializableSchema | JsonSchemaFormat)[]
    options?: ToolStrategyOptions
): ToolStrategy<StructuredResponseT>

schema 必填

定义结构化输出格式的 Schema。支持：

  • Zod Schema：Zod schema
  • Standard Schema：任何实现 Standard Schema 规范的 schema
  • JSON Schema：JSON Schema 对象

options.toolMessageContent

生成结构化输出时返回的工具消息自定义内容。 如果未提供，默认显示包含结构化响应数据的消息。

options.handleError

用于自定义错误处理策略的可选 handleError 参数所在的选项对象。

  • true：使用默认错误模板捕获所有错误（默认）
  • False：不进行重试，让异常向上抛出
  • (error: ToolStrategyError) => string | Promise<string>：使用提供的消息进行重试或抛出错误
ts
import * as z from "zod";
import { createAgent, toolStrategy } from "langchain";

const ProductReview = z.object({
    rating: z.number().min(1).max(5).optional(),
    sentiment: z.enum(["positive", "negative"]),
    keyPoints: z.array(z.string()).describe("The key points of the review. Lowercase, 1-3 words each."),
});

const agent = createAgent({
    model: "gpt-5.5",
    tools: [],
    responseFormat: toolStrategy(ProductReview)
})

const result = await agent.invoke({
    "messages": [{"role": "user", "content": "Analyze this review: 'Great product: 5 out of 5 stars. Fast shipping, but expensive'"}]
})

console.log(result.structuredResponse);
// { "rating": 5, "sentiment": "positive", "keyPoints": ["fast shipping", "expensive"] }
ts
import * as v from "valibot";
import { toStandardJsonSchema } from "@valibot/to-json-schema";
import { createAgent, toolStrategy } from "langchain";

const ProductReview = toStandardJsonSchema(
    v.object({
        rating: v.optional(v.pipe(v.number(), v.minValue(1), v.maxValue(5))),
        sentiment: v.picklist(["positive", "negative"]),
        keyPoints: v.pipe(v.array(v.string()), v.description("The key points of the review. Lowercase, 1-3 words each.")),
    })
);

const agent = createAgent({
    model: "gpt-5.5",
    tools: [],
    responseFormat: toolStrategy(ProductReview)
})

const result = await agent.invoke({
    messages: [{"role": "user", "content": "Analyze this review: 'Great product: 5 out of 5 stars. Fast shipping, but expensive'"}]
})

console.log(result.structuredResponse);
// { "rating": 5, "sentiment": "positive", "keyPoints": ["fast shipping", "expensive"] }
ts
import { createAgent, toolStrategy } from "langchain";

const productReviewSchema = {
    "type": "object",
    "description": "Analysis of a product review.",
    "properties": {
        "rating": {
            "type": ["integer", "null"],
            "description": "The rating of the product (1-5)",
            "minimum": 1,
            "maximum": 5
        },
        "sentiment": {
            "type": "string",
            "enum": ["positive", "negative"],
            "description": "The sentiment of the review"
        },
        "key_points": {
            "type": "array",
            "items": {"type": "string"},
            "description": "The key points of the review"
        }
    },
    "required": ["sentiment", "key_points"]
}

const agent = createAgent({
    model: "gpt-5.5",
    tools: [],
    responseFormat: toolStrategy(productReviewSchema)
});

const result = await agent.invoke({
    messages: [{"role": "user", "content": "Analyze this review: 'Great product: 5 out of 5 stars. Fast shipping, but expensive'"}]
})

console.log(result.structuredResponse);
// { "rating": 5, "sentiment": "positive", "keyPoints": ["fast shipping", "expensive"] }
ts
import * as z from "zod";
import { createAgent, toolStrategy } from "langchain";

const ProductReview = z.object({
    rating: z.number().min(1).max(5).optional(),
    sentiment: z.enum(["positive", "negative"]),
    keyPoints: z.array(z.string()).describe("The key points of the review. Lowercase, 1-3 words each."),
});

const CustomerComplaint = z.object({
    issueType: z.enum(["product", "service", "shipping", "billing"]),
    severity: z.enum(["low", "medium", "high"]),
    description: z.string().describe("Brief description of the complaint"),
});

const agent = createAgent({
    model: "gpt-5.5",
    tools: [],
    responseFormat: toolStrategy([ProductReview, CustomerComplaint])
});

const result = await agent.invoke({
    messages: [{"role": "user", "content": "Analyze this review: 'Great product: 5 out of 5 stars. Fast shipping, but expensive'"}]
})

console.log(result.structuredResponse);
// { "rating": 5, "sentiment": "positive", "keyPoints": ["fast shipping", "expensive"] }

自定义工具消息内容 (Custom tool message content) ​

toolMessageContent 参数允许你自定义在生成结构化输出时出现在对话历史记录中的消息：

ts
import * as z from "zod";
import { createAgent, toolStrategy } from "langchain";

const MeetingAction = z.object({
    task: z.string().describe("The specific task to be completed"),
    assignee: z.string().describe("Person responsible for the task"),
    priority: z.enum(["low", "medium", "high"]).describe("Priority level"),
});

const agent = createAgent({
    model: "gpt-5.5",
    tools: [],
    responseFormat: toolStrategy(MeetingAction, {
        toolMessageContent: "Action item captured and added to meeting notes!"
    })
});

const result = await agent.invoke({
    messages: [{"role": "user", "content": "From our meeting: Sarah needs to update the project timeline as soon as possible"}]
});

console.log(result);
/**
 * {
 *   messages: [
 *     { role: "user", content: "From our meeting: Sarah needs to update the project timeline as soon as possible" },
 *     { role: "assistant", content: "Action item captured and added to meeting notes!", tool_calls: [ { name: "MeetingAction", args: { task: "update the project timeline", assignee: "Sarah", priority: "high" }, id: "call_456" } ] },
 *     { role: "tool", content: "Action item captured and added to meeting notes!", tool_call_id: "call_456", name: "MeetingAction" }
 *   ],
 *   structuredResponse: { task: "update the project timeline", assignee: "Sarah", priority: "high" }
 * }
 */

若不使用 toolMessageContent，我们会看到：

ts
# console.log(result);
/**
 * {
 *   messages: [
 *     ...
 *     { role: "tool", content: "Returning structured response: {'task': 'update the project timeline', 'assignee': 'Sarah', 'priority': 'high'}", tool_call_id: "call_456", name: "MeetingAction" }
 *   ],
 *   structuredResponse: { task: "update the project timeline", assignee: "Sarah", priority: "high" }
 * }
 */

错误处理 (Error handling) ​

模型通过工具调用生成结构化输出时可能会犯错。LangChain 提供了智能重试机制来自动处理这些错误。

多个结构化输出错误 (Multiple structured outputs error) ​

当模型错误地调用了多个结构化输出工具时，智能体会通过 ToolMessage 提供错误反馈，并提示模型进行重试：

ts
import * as z from "zod";
import { createAgent, toolStrategy } from "langchain";

const ContactInfo = z.object({
    name: z.string().describe("Person's name"),
    email: z.string().describe("Email address"),
});

const EventDetails = z.object({
    event_name: z.string().describe("Name of the event"),
    date: z.string().describe("Event date"),
});

const agent = createAgent({
    model: "gpt-5.5",
    tools: [],
    responseFormat: toolStrategy([ContactInfo, EventDetails]),
});

const result = await agent.invoke({
    messages: [
        {
        role: "user",
        content:
            "Extract info: John Doe (john@email.com) is organizing Tech Conference on March 15th",
        },
    ],
});

console.log(result);

/**
 * {
 *   messages: [
 *     { role: "user", content: "Extract info: John Doe (john@email.com) is organizing Tech Conference on March 15th" },
 *     { role: "assistant", content: "", tool_calls: [ { name: "ContactInfo", args: { name: "John Doe", email: "john@email.com" }, id: "call_1" }, { name: "EventDetails", args: { event_name: "Tech Conference", date: "March 15th" }, id: "call_2" } ] },
 *     { role: "tool", content: "Error: Model incorrectly returned multiple structured responses (ContactInfo, EventDetails) when only one is expected.\n Please fix your mistakes.", tool_call_id: "call_1", name: "ContactInfo" },
 *     { role: "tool", content: "Error: Model incorrectly returned multiple structured responses (ContactInfo, EventDetails) when only one is expected.\n Please fix your mistakes.", tool_call_id: "call_2", name: "EventDetails" },
 *     { role: "assistant", content: "", tool_calls: [ { name: "ContactInfo", args: { name: "John Doe", email: "john@email.com" }, id: "call_3" } ] },
 *     { role: "tool", content: "Returning structured response: {'name': 'John Doe', 'email': 'john@email.com'}", tool_call_id: "call_3", name: "ContactInfo" }
 *   ],
 *   structuredResponse: { name: "John Doe", email: "john@email.com" }
 * }
 */

Schema 校验错误 (Schema validation error) ​

当结构化输出不匹配期望的 Schema 时，智能体会提供具体的错误反馈：

ts
import * as z from "zod";
import { createAgent, toolStrategy } from "langchain";

const ProductRating = z.object({
    rating: z.number().min(1).max(5).describe("Rating from 1-5"),
    comment: z.string().describe("Review comment"),
});

const agent = createAgent({
    model: "gpt-5.5",
    tools: [],
    responseFormat: toolStrategy(ProductRating),
});

const result = await agent.invoke({
    messages: [
        {
        role: "user",
        content: "Parse this: Amazing product, 10/10!",
        },
    ],
});

console.log(result);

/**
 * {
 *   messages: [
 *     { role: "user", content: "Parse this: Amazing product, 10/10!" },
 *     { role: "assistant", content: "", tool_calls: [ { name: "ProductRating", args: { rating: 10, comment: "Amazing product" }, id: "call_1" } ] },
 *     { role: "tool", content: "Error: Failed to parse structured output for tool 'ProductRating': 1 validation error for ProductRating\nrating\n  Input should be less than or equal to 5 [type=less_than_equal, input_value=10, input_type=int].\n Please fix your mistakes.", tool_call_id: "call_1", name: "ProductRating" },
 *     { role: "assistant", content: "", tool_calls: [ { name: "ProductRating", args: { rating: 5, comment: "Amazing product" }, id: "call_2" } ] },
 *     { role: "tool", content: "Returning structured response: {'rating': 5, 'comment': 'Amazing product'}", tool_call_id: "call_2", name: "ProductRating" }
 *   ],
 *   structuredResponse: { rating: 5, comment: "Amazing product" }
 * }
 */

错误处理策略 (Error handling strategies) ​

你可以使用 handleErrors 参数自定义错误的具体处理方式：

自定义错误消息：

ts
const responseFormat = toolStrategy(ProductRating, {
    handleError: "Please provide a valid rating between 1-5 and include a comment."
)

// 错误消息变为：
// { role: "tool", content: "Please provide a valid rating between 1-5 and include a comment." }

仅处理特定异常：

ts
import { ToolInputParsingException } from "@langchain/core/tools";

const responseFormat = toolStrategy(ProductRating, {
    handleError: (error: ToolStrategyError) => {
        if (error instanceof ToolInputParsingException) {
        return "Please provide a valid rating between 1-5 and include a comment.";
        }
        return error.message;
    }
)

// 仅校验错误会使用默认消息重试：
// { role: "tool", content: "Error: Failed to parse structured output for tool 'ProductRating': ...\n Please fix your mistakes." }

处理多种异常类型：

ts
const responseFormat = toolStrategy(ProductRating, {
    handleError: (error: ToolStrategyError) => {
        if (error instanceof ToolInputParsingException) {
        return "Please provide a valid rating between 1-5 and include a comment.";
        }
        if (error instanceof CustomUserError) {
        return "This is a custom user error.";
        }
        return error.message;
    }
)

不进行错误处理：

ts
const responseFormat = toolStrategy(ProductRating, {
    handleError: false  // 抛出所有错误
)