ChatOpenAI 集成 - 以编程方式使用文档

OpenAI 是一家人工智能（AI）研究实验室。

本指南将帮助您开始使用 OpenAI 聊天模型。有关所有 ChatOpenAI 功能和配置的详细文档，请访问 API 参考.

概述

集成详情

类	包	可序列化	PY 支持	下载量	版本
`ChatOpenAI`	`@langchain/openai`	✅	✅	!NPM - 下载量	!NPM - 版本

模型功能

请参阅下表标题中的链接，了解如何使用特定功能。

工具调用	结构化输出	图像输入	音频输入	视频输入	Token 级流式输出	Token 使用量	对数概率
✅	✅	✅	❌	❌	✅	✅	✅

设置

要访问 OpenAI 聊天模型，您需要创建一个 OpenAI 账户，获取一个 API 密钥，并安装 @langchain/openai 集成包。

凭据

访问 OpenAI 网站注册 OpenAI 并生成 API 密钥。完成后，设置 OPENAI_API_KEY 环境变量：

如果您想获取模型调用的自动跟踪，还可以设置您的 LangSmith API 密钥，方法：取消下方注释：

# export LANGSMITH_TRACING="true"
# export LANGSMITH_API_KEY="your-api-key"

安装

LangChain ChatOpenAI 集成位于 @langchain/openai package:

npm install @langchain/openai @langchain/core

yarn add @langchain/openai @langchain/core

pnpm add @langchain/openai @langchain/core

实例化

现在我们可以实例化模型对象并生成聊天补全：

const llm = new ChatOpenAI({
  model: "gpt-5.5",
  temperature: 0,
  // other params...
})

调用

const aiMsg = await llm.invoke([
  {
    role: "system",
    content: "You are a helpful assistant that translates English to French. Translate the user sentence.",
  },
  {
    role: "user",
    content: "I love programming."
  },
])
aiMsg

AIMessage {
  "id": "chatcmpl-ADItECqSPuuEuBHHPjeCkh9wIO1H5",
  "content": "J'adore la programmation.",
  "additional_kwargs": {},
  "response_metadata": {
    "tokenUsage": {
      "completionTokens": 5,
      "promptTokens": 31,
      "totalTokens": 36
    },
    "finish_reason": "stop",
    "system_fingerprint": "fp_5796ac6771"
  },
  "tool_calls": [],
  "invalid_tool_calls": [],
  "usage_metadata": {
    "input_tokens": 31,
    "output_tokens": 5,
    "total_tokens": 36
  }
}

console.log(aiMsg.content)

J'adore la programmation.

自定义 URL

您可以通过传递一个来自定义 SDK 发送请求的基础 URL configuration 参数，如下所示：

const llmWithCustomURL = new ChatOpenAI({
  model: "gpt-5.5",
  temperature: 0.9,
  configuration: {
    baseURL: "https://your_custom_url.com",
  },
});

await llmWithCustomURL.invoke("Hi there!");

这个 configuration 字段也接受其他 ClientOptions 官方 SDK 接受的参数。

如果您托管在 Azure OpenAI 上，请参阅专用页面.

自定义请求头

您可以用相同的方式指定自定义请求头 configuration field:

const llmWithCustomHeaders = new ChatOpenAI({
  model: "gpt-5.5",
  temperature: 0.9,
  configuration: {
    defaultHeaders: {
      "Authorization": `Bearer SOME_CUSTOM_VALUE`,
    },
  },
});

await llmWithCustomHeaders.invoke("Hi there!");

禁用流式使用元数据

一些代理或第三方提供商提供的 API 接口与 OpenAI 大致相同，但不支持最近添加的 stream_options 参数来返回流式使用。您可以使用 ChatOpenAI 来访问这些提供商，方法是通过禁用流式使用，如下所示：

const llmWithoutStreamUsage = new ChatOpenAI({
  model: "gpt-5.5",
  temperature: 0.9,
  streamUsage: false,
  configuration: {
    baseURL: "https://proxy.com",
  },
});

await llmWithoutStreamUsage.invoke("Hi there!");

调用微调模型

您可以通过传入相应的 modelName parameter.

这通常采用以下形式 ft:{OPENAI_MODEL_NAME}:{ORG_NAME}::{MODEL_ID}。例如：

const fineTunedLlm = new ChatOpenAI({
  temperature: 0.9,
  model: "ft:gpt-3.5-turbo-0613:{ORG_NAME}::{MODEL_ID}",
});

await fineTunedLlm.invoke("Hi there!");

生成元数据

如果您需要 logprobs 或 token 使用情况等额外信息，这些信息将直接在 invoke 响应中的 response_metadata 消息字段中返回。

// See https://cookbook.openai.com/examples/using_logprobs for details
const llmWithLogprobs = new ChatOpenAI({
  model: "gpt-5.5",
  logprobs: true,
  // topLogprobs: 5,
});

const responseMessageWithLogprobs = await llmWithLogprobs.invoke("Hi there!");
console.dir(responseMessageWithLogprobs.response_metadata.logprobs, { depth: null });

{
  content: [
    {
      token: 'Hello',
      logprob: -0.0004740447,
      bytes: [ 72, 101, 108, 108, 111 ],
      top_logprobs: []
    },
    {
      token: '!',
      logprob: -0.00004334534,
      bytes: [ 33 ],
      top_logprobs: []
    },
    {
      token: ' How',
      logprob: -0.000030113732,
      bytes: [ 32, 72, 111, 119 ],
      top_logprobs: []
    },
    {
      token: ' can',
      logprob: -0.0004797665,
      bytes: [ 32, 99, 97, 110 ],
      top_logprobs: []
    },
    {
      token: ' I',
      logprob: -7.89631e-7,
      bytes: [ 32, 73 ],
      top_logprobs: []
    },
    {
      token: ' assist',
      logprob: -0.114006,
      bytes: [
         32,  97, 115,
        115, 105, 115,
        116
      ],
      top_logprobs: []
    },
    {
      token: ' you',
      logprob: -4.3202e-7,
      bytes: [ 32, 121, 111, 117 ],
      top_logprobs: []
    },
    {
      token: ' today',
      logprob: -0.00004501419,
      bytes: [ 32, 116, 111, 100, 97, 121 ],
      top_logprobs: []
    },
    {
      token: '?',
      logprob: -0.000010206721,
      bytes: [ 63 ],
      top_logprobs: []
    }
  ],
  refusal: null
}

自定义工具

自定义工具支持具有任意字符串输入的工具。当您期望字符串参数较长或复杂时，它们特别有用。

如果您使用的模型支持自定义工具，您可以使用 ChatOpenAI 类和 customTool 函数来创建自定义工具。

const codeTool = customTool(
  async () => {
    // ... Add code to execute the input
    return "Code executed successfully";
  },
  {
    name: "execute_code",
    description: "Execute a code snippet",
    format: { type: "text" },
  }
);

const model = new ChatOpenAI({ model: "gpt-5.5" });

const agent = createAgent({
  model,
  tools: [codeTool],
});

const result = await agent.invoke({
  messages: [new HumanMessage("Use the tool to execute the code")],
});
console.log(result);

上下文无关语法

OpenAI 支持指定上下文无关语法用于自定义工具输入 lark or regex 格式。参见 OpenAI 文档了解更多详情。 format 参数可以传入 customTool ，如下所示：

const MATH_GRAMMAR = `
start: expr
expr: term (SP ADD SP term)* -> add
| term
term: factor (SP MUL SP factor)* -> mul
| factor
factor: INT
SP: \" \"
ADD: \"+\"
MUL: \"*\"
%import common.INT
`;

const doMath = customTool(
  async () => {
    // ... Add code to parse and execute the input
    return "27";
  },
  {
    name: "do_math",
    description: "Evaluate a math expression",
    format: { type: "grammar", definition: MATH_GRAMMAR, syntax: "lark" },
  }
);

const model = new ChatOpenAI({ model: "gpt-5.5" });

const agent = createAgent({
  model,
  tools: [doMath],
});

const result = await agent.invoke({
  messages: [new HumanMessage("Use the tool to calculate 3^3")],
});
console.log(result);

`strict: true`

截至 2024 年 8 月 6 日，OpenAI 在调用工具时支持 strict 参数，该参数将强制模型遵守工具参数模式。查看更多.

以下是使用工具调用的示例。向 strict: true 传递额外的 .bindTools 参数会将该参数传递给所有工具定义：

const weatherTool = tool((_) => "no-op", {
  name: "get_current_weather",
  description: "Get the current weather",
  schema: z.object({
    location: z.string(),
  }),
})

const llmWithStrictTrue = new ChatOpenAI({
  model: "gpt-5.5",
}).bindTools([weatherTool], {
  strict: true,
  tool_choice: weatherTool.name,
});

// Although the question is not about the weather, it will call the tool with the correct arguments
// because we passed `tool_choice` and `strict: true`.
const strictTrueResult = await llmWithStrictTrue.invoke("What is 127862 times 12898 divided by 2?");

console.dir(strictTrueResult.tool_calls, { depth: null });

[
  {
    name: 'get_current_weather',
    args: { location: 'current' },
    type: 'tool_call',
    id: 'call_hVFyYNRwc6CoTgr9AQFQVjm9'
  }
]

如果您只想将此参数应用于部分工具，也可以直接传递 OpenAI 格式的工具 schema：

const toolSchema = {
  type: "function",
  function: {
    name: "get_current_weather",
    description: "Get the current weather",
    strict: true,
    parameters: zodToJsonSchema(
      z.object({
        location: z.string(),
      })
    ),
  },
};

const llmWithStrictTrueTools = new ChatOpenAI({
  model: "gpt-5.5",
}).bindTools([toolSchema], {
  strict: true,
});

const weatherToolResult = await llmWithStrictTrueTools.invoke([{
  role: "user",
  content: "What is the current weather in London?"
}])

weatherToolResult.tool_calls;

[
  {
    name: 'get_current_weather',
    args: { location: 'London' },
    type: 'tool_call',
    id: 'call_EOSejtax8aYtqpchY8n8O82l'
  }
]

结构化输出

我们还可以传递 strict: true 到 .withStructuredOutput()。以下是示例：

const traitSchema = z.object({
  traits: z.array(z.string()).describe("A list of traits contained in the input"),
});

const structuredLlm = new ChatOpenAI({
  model: "gpt-5.4-mini",
}).withStructuredOutput(traitSchema, {
  name: "extract_traits",
  strict: true,
});

await structuredLlm.invoke([{
  role: "user",
  content: `I am 6'5" tall and love fruit.`
}]);

{ traits: [ `6'5" tall`, 'love fruit' ] }

Responses API

OpenAI 支持一个 Responses API，面向构建智能体应用。它包含一套内置工具，包括网络和文件搜索。它还支持管理对话状态，允许您继续对话线程而无需明确传入之前的消息。

ChatOpenAI 如果使用了这些功能之一，将路由到 Responses API。您也可以指定 useResponsesApi: true 在实例化时 ChatOpenAI.

内置工具

为 ChatOpenAI 配备内置工具将以其外部信息（如文件或网络中的上下文）为依据进行响应。The AIMessage 从模型生成的内容将包含有关内置工具调用的信息。

网络搜索

要触发网络搜索，请传递 {"type": "web_search_preview"} 给模型，就像使用其他工具一样。

建议

您也可以将内置工具作为调用参数传递：

llm.invoke("...", { tools: [{ type: "web_search_preview" }] });

const llm = new ChatOpenAI({ model: "gpt-5.4-mini" }).bindTools([
  { type: "web_search_preview" },
]);

await llm.invoke("What was a positive news story from today?");

请注意，响应包含结构化的内容块，包括响应的文本和 OpenAI 注释引用其来源。输出消息还将包含来自任何工具调用的信息。

文件搜索

要触发文件搜索，请传递一个文件搜索工具给模型，就像使用其他工具一样。您需要填充一个 OpenAI 管理的向量存储，并在工具定义中包含向量存储 ID。参见 OpenAI 文档了解更多详情。

const llm = new ChatOpenAI({ model: "gpt-5.4-mini" }).bindTools([
  { type: "file_search", vector_store_ids: ["vs..."] },
]);

await llm.invoke("Is deep research by OpenAI?");

与网络搜索一样，响应将包含带有引用的内容块。它还将包含来自内置工具调用的信息。

计算机使用

ChatOpenAI 支持 computer-use-preview 模型，这是一个用于内置计算机使用工具的专用模型。要启用，请传递一个计算机使用工具就像传递其他工具一样。

目前计算机使用的工具输出出现在 AIMessage.additional_kwargs.tool_outputs中。要回复计算机使用工具调用，您需要设置 additional_kwargs.type: "computer_call_output" 同时创建一个相应的 ToolMessage.

参见 OpenAI 文档了解更多详情。

const findComputerCall = (message: AIMessage) => {
  const toolOutputs = message.additional_kwargs.tool_outputs as
    | { type: "computer_call"; call_id: string; action: { type: string } }[]
    | undefined;

  return toolOutputs?.find((toolOutput) => toolOutput.type === "computer_call");
};

const llm = new ChatOpenAI({ model: "computer-use-preview" })
  .bindTools([
    {
      type: "computer-preview",
      display_width: 1024,
      display_height: 768,
      environment: "browser",
    },
  ])
  .bind({ truncation: "auto" });

let message = await llm.invoke("Check the latest OpenAI news on bing.com.");
const computerCall = findComputerCall(message);

if (computerCall) {
  // Act on a computer call action
  const screenshot = await fs.readFile("./screenshot.png", {
    encoding: "base64",
  });

  message = await llm.invoke(
    [
      new ToolMessage({
        additional_kwargs: { type: "computer_call_output" },
        tool_call_id: computerCall.call_id,
        content: [
          {
            type: "computer_screenshot",
            image_url: `data:image/png;base64,${screenshot}`,
          },
        ],
      }),
    ],
    { previous_response_id: message.response_metadata["id"] }
  );
}

代码解释器

ChatOpenAI 允许您使用内置代码解释器工具以支持沙箱化的代码生成和执行。

const llm = new ChatOpenAI({
  model: "o4-mini",
  useResponsesApi: true,
});

const llmWithTools = llm.bindToools([
  {
    type: "code_interpreter",
    // Creates a new container
    container: { type: "auto" }
  },
]);

const response = await llmWithTools.invoke(
  "Write and run code to answer the question: what is 3^3?"
);

请注意，上面的命令会创建一个新的容器。我们可以通过指定现有容器 ID 来跨调用重用容器。

const tool_outputs: Record<string, any>[] = response.additional_kwargs.tool_outputs
const container_id = tool_outputs[0].container_id

const llmWithTools = llm.bindTools([
  {
    type: "code_interpreter",
    // Reuses container from the last call
    container: container_id,
  },
]);

远程 MCP

ChatOpenAI 支持内置的远程 MCP 工具，允许模型生成的对 MCP 服务器的调用在 OpenAI 服务器上执行。

const llm = new ChatOpenAI({
    model: "o4-mini",
    useResponsesApi: true,
});

const llmWithMcp = llm.bindTools([
    {
        type: "mcp",
        server_label: "deepwiki",
        server_url: "https://mcp.deepwiki.com/mcp",
        require_approval: "never"
    }
]);

const response = await llmWithMcp.invoke(
    "What transport protocols does the 2025-03-26 version of the MCP spec (modelcontextprotocol/modelcontextprotocol) support?"
);

提示

MCP 审批

在收到指示时，OpenAI 会在调用远程 MCP 服务器之前请求审批。

在上面的命令中，我们指示模型永远不需要审批。我们也可以将模型配置为始终请求审批，或始终对特定工具请求审批：

...
const llmWithMcp = llm.bindTools([
  {
    type: "mcp",
    server_label: "deepwiki",
    server_url: "https://mcp.deepwiki.com/mcp",
    require_approval: {
      always: {
        tool_names: ["read_wiki_structure"],
      },
    },
  },
]);
const response = await llmWithMcp.invoke(
    "What transport protocols does the 2025-03-26 version of the MCP spec (modelcontextprotocol/modelcontextprotocol) support?"
);

使用此配置，响应可以包含类型为 mcp_approval_request的工具输出。要提交审批请求的审批，您可以将其构建为后续消息中的内容块：

const approvals = [];
if (Array.isArray(response.additional_kwargs.tool_outputs)) {
  for (const content of response.additional_kwargs.tool_outputs) {
    if (content.type === "mcp_approval_request") {
      approvals.push({
        type: "mcp_approval_response",
        approval_request_id: content.id,
        approve: true,
      });
    }
  }
}

const nextResponse = await model.invoke(
  [
    response,
    new HumanMessage({ content: approvals }),
  ],
);

图像生成

ChatOpenAI 允许您使用内置的图像生成工具通过 responses API 在多轮对话中创建图像。

const llm = new ChatOpenAI({
  model: "gpt-5.5",
  useResponsesApi: true,
});

const llmWithImageGeneration = llm.bindTools([
  {
    type: "image_generation",
    quality: "low",
  }
]);

const response = await llmWithImageGeneration.invoke(
  "Draw a random short word in green font."
)

推理模型

当使用推理模型（如 o1）时，结构化输出的默认方法是 withStructuredOutput OpenAI 内置的结构化输出方法（相当于传入 method: "jsonSchema" 作为 withStructuredOutput的选项）。JSON schema 与其他模型基本相同，但有一个重要的注意事项：在定义 schema 时， z.optional() 不会被尊重，您应该改用 z.nullable().

示例如下：

// Will not work
const reasoningModelSchemaOptional = z.object({
  color: z.optional(z.string()).describe("A color mentioned in the input"),
});

const reasoningModelOptionalSchema = new ChatOpenAI({
  model: "o1",
}).withStructuredOutput(reasoningModelSchemaOptional, {
  name: "extract_color",
});

await reasoningModelOptionalSchema.invoke([{
  role: "user",
  content: `I am 6'5" tall and love fruit.`
}]);

{ color: 'No color mentioned' }

以下是配合 z.nullable():

// Will not work
const reasoningModelSchemaNullable = z.object({
  color: z.nullable(z.string()).describe("A color mentioned in the input"),
});

const reasoningModelNullableSchema = new ChatOpenAI({
  model: "o1",
}).withStructuredOutput(reasoningModelSchemaNullable, {
  name: "extract_color",
});

await reasoningModelNullableSchema.invoke([{
  role: "user",
  content: `I am 6'5" tall and love fruit.`
}]);

{ color: null }

提示缓存

较新的 OpenAI 模型会自动缓存提示的部分内容（如果您的输入超过一定大小，截至撰写本文时为 1024 个 tokens），以降低需要长上下文的用例的成本。

Note: 为给定查询缓存的 token 数量尚未在 AIMessage.usage_metadata中标准化，而是包含在 AIMessage.response_metadata field.

中的示例

// @lc-docs-hide-cell

const CACHED_TEXT = `## Components

LangChain provides standard, extendable interfaces and external integrations for various components useful for building with LLMs.
Some components LangChain implements, some components we rely on third-party integrations for, and others are a mix.

### Chat models

<span data-heading-keywords="chat model,chat models"></span>

Language models that use a sequence of messages as inputs and return chat messages as outputs (as opposed to using plain text).
These are generally newer models (older models are generally \`LLMs\`, see below).
Chat models support the assignment of distinct roles to conversation messages, helping to distinguish messages from the AI, users, and instructions such as system messages.

Although the underlying models are messages in, message out, the LangChain wrappers also allow these models to take a string as input.
This gives them the same interface as LLMs (and simpler to use).
When a string is passed in as input, it will be converted to a \`HumanMessage\` under the hood before being passed to the underlying model.

LangChain does not host any Chat Models, rather we rely on third party integrations.

We have some standardized parameters when constructing ChatModels:

- \`model\`: the name of the model

Chat Models also accept other parameters that are specific to that integration.


<aside class="callout"><strong>警告</strong>

**Some chat models have been fine-tuned for **tool calling** and provide a dedicated API for it.**

Generally, such models are better at tool calling than non-fine-tuned models, and are recommended for use cases that require tool calling.
Please see the [tool calling section](/oss/javascript/langchain/tools.html) for more information.

</aside>


For specifics on how to use chat models, see the [relevant how-to guides here](/oss/javascript/langchain/models.html).

#### Multimodality

Some chat models are multimodal, accepting images, audio and even video as inputs.
These are still less common, meaning model providers haven't standardized on the "best" way to define the API.
Multimodal outputs are even less common. As such, we've kept our multimodal abstractions fairly light weight
and plan to further solidify the multimodal APIs and interaction patterns as the field matures.

In LangChain, most chat models that support multimodal inputs also accept those values in OpenAI's content blocks format.
So far this is restricted to image inputs. For models like Gemini which support video and other bytes input, the APIs also support the native, model-specific representations.

For specifics on how to use multimodal models, see the [relevant how-to guides here](/oss/javascript/how-to/#multimodal).

### LLMs

<span data-heading-keywords="llm,llms"></span>


<aside class="callout"><strong>警告</strong>

**Pure text-in/text-out LLMs tend to be older or lower-level. Many popular models are best used as [chat completion models](/oss/javascript/langchain/models.html),**

even for non-chat use cases.

You are probably looking for [the section above instead](/oss/javascript/langchain/models.html).

</aside>


Language models that takes a string as input and returns a string.
These are traditionally older models (newer models generally are [Chat Models](/oss/javascript/langchain/models.html), see above).

Although the underlying models are string in, string out, the LangChain wrappers also allow these models to take messages as input.
This gives them the same interface as [Chat Models](/oss/javascript/langchain/models.html).
When messages are passed in as input, they will be formatted into a string under the hood before being passed to the underlying model.

LangChain does not host any LLMs, rather we rely on third party integrations.

For specifics on how to use LLMs, see the [relevant how-to guides here](/oss/javascript/langchain/models.html).

### Message types

Some language models take an array of messages as input and return a message.
There are a few different types of messages.
All messages have a \`role\`, \`content\`, and \`response_metadata\` property.

The \`role\` describes WHO is saying the message.
LangChain has different message classes for different roles.

The \`content\` property describes the content of the message.
This can be a few different things:

- A string (most models deal this type of content)
- A List of objects (this is used for multi-modal input, where the object contains information about that input type and that input location)

#### HumanMessage

This represents a message from the user.

#### AIMessage

This represents a message from the model. In addition to the \`content\` property, these messages also have:

**\`response_metadata\`**

The \`response_metadata\` property contains additional metadata about the response. The data here is often specific to each model provider.
This is where information like log-probs and token usage may be stored.

**\`tool_calls\`**

These represent a decision from an language model to call a tool. They are included as part of an \`AIMessage\` output.
They can be accessed from there with the \`.tool_calls\` property.

This property returns a list of \`ToolCall\`s. A \`ToolCall\` is an object with the following arguments:

- \`name\`: The name of the tool that should be called.
- \`args\`: The arguments to that tool.
- \`id\`: The id of that tool call.

#### SystemMessage

This represents a system message, which tells the model how to behave. Not every model provider supports this.

#### ToolMessage

This represents the result of a tool call. In addition to \`role\` and \`content\`, this message has:

- a \`tool_call_id\` field which conveys the id of the call to the tool that was called to produce this result.
- an \`artifact\` field which can be used to pass along arbitrary artifacts of the tool execution which are useful to track but which should not be sent to the model.

#### (Legacy) FunctionMessage

This is a legacy message type, corresponding to OpenAI's legacy function-calling API. \`ToolMessage\` should be used instead to correspond to the updated tool-calling API.

This represents the result of a function call. In addition to \`role\` and \`content\`, this message has a \`name\` parameter which conveys the name of the function that was called to produce this result.

### Prompt templates

<span data-heading-keywords="prompt,prompttemplate,chatprompttemplate"></span>

Prompt templates help to translate user input and parameters into instructions for a language model.
This can be used to guide a model's response, helping it understand the context and generate relevant and coherent language-based output.

Prompt Templates take as input an object, where each key represents a variable in the prompt template to fill in.

Prompt Templates output a PromptValue. This PromptValue can be passed to an LLM or a ChatModel, and can also be cast to a string or an array of messages.
The reason this PromptValue exists is to make it easy to switch between strings and messages.

There are a few different types of prompt templates:

#### String PromptTemplates

These prompt templates are used to format a single string, and generally are used for simpler inputs.
For example, a common way to construct and use a PromptTemplate is as follows:

\`\`\`typescript


const promptTemplate = PromptTemplate.fromTemplate(
  "Tell me a joke about {topic}"
);

await promptTemplate.invoke({ topic: "cats" });
\`\`\`

#### ChatPromptTemplates

These prompt templates are used to format an array of messages. These "templates" consist of an array of templates themselves.
For example, a common way to construct and use a ChatPromptTemplate is as follows:

\`\`\`typescript


const promptTemplate = ChatPromptTemplate.fromMessages([
  ["system", "You are a helpful assistant"],
  ["user", "Tell me a joke about {topic}"],
]);

await promptTemplate.invoke({ topic: "cats" });
\`\`\`

In the above example, this ChatPromptTemplate will construct two messages when called.
The first is a system message, that has no variables to format.
The second is a HumanMessage, and will be formatted by the \`topic\` variable the user passes in.

#### MessagesPlaceholder

<span data-heading-keywords="messagesplaceholder"></span>

This prompt template is responsible for adding an array of messages in a particular place.
In the above ChatPromptTemplate, we saw how we could format two messages, each one a string.
But what if we wanted the user to pass in an array of messages that we would slot into a particular spot?
This is how you use MessagesPlaceholder.

\`\`\`typescript

  ChatPromptTemplate,
  MessagesPlaceholder,
} from "@langchain/core/prompts";


const promptTemplate = ChatPromptTemplate.fromMessages([
  ["system", "You are a helpful assistant"],
  new MessagesPlaceholder("msgs"),
]);

promptTemplate.invoke({ msgs: [new HumanMessage({ content: "hi!" })] });
\`\`\`

This will produce an array of two messages, the first one being a system message, and the second one being the HumanMessage we passed in.
If we had passed in 5 messages, then it would have produced 6 messages in total (the system message plus the 5 passed in).
This is useful for letting an array of messages be slotted into a particular spot.

An alternative way to accomplish the same thing without using the \`MessagesPlaceholder\` class explicitly is:

\`\`\`typescript
const promptTemplate = ChatPromptTemplate.fromMessages([
  ["system", "You are a helpful assistant"],
  ["placeholder", "{msgs}"], // <-- This is the changed part
]);
\`\`\`

For specifics on how to use prompt templates, see the [relevant how-to guides here](/oss/javascript/how-to/#prompt-templates).

### Example Selectors

One common prompting technique for achieving better performance is to include examples as part of the prompt.
This gives the language model concrete examples of how it should behave.
Sometimes these examples are hardcoded into the prompt, but for more advanced situations it may be nice to dynamically select them.
Example Selectors are classes responsible for selecting and then formatting examples into prompts.

For specifics on how to use example selectors, see the [relevant how-to guides here](/oss/javascript/how-to/#example-selectors).

### Output parsers

<span data-heading-keywords="output parser"></span>


<aside class="callout"><strong>提示</strong>

**The information here refers to parsers that take a text output from a model try to parse it into a more structured representation.**

More and more models are supporting function (or tool) calling, which handles this automatically.
It is recommended to use function/tool calling rather than output parsing.
See the [LangChain tools documentation](/oss/javascript/langchain/tools.html).


</aside>


Responsible for taking the output of a model and transforming it to a more suitable format for downstream tasks.
Useful when you are using LLMs to generate structured data, or to normalize output from chat models and LLMs.

There are two main methods an output parser must implement:

- "Get format instructions": A method which returns a string containing instructions for how the output of a language model should be formatted.
- "Parse": A method which takes in a string (assumed to be the response from a language model) and parses it into some structure.

And then one optional one:

- "Parse with prompt": A method which takes in a string (assumed to be the response from a language model) and a prompt (assumed to be the prompt that generated such a response) and parses it into some structure. The prompt is largely provided in the event the OutputParser wants to retry or fix the output in some way, and needs information from the prompt to do so.

Output parsers accept a string or \`BaseMessage\` as input and can return an arbitrary type.

LangChain has many different types of output parsers. This is a list of output parsers LangChain supports. The table below has various pieces of information:

**Name**: The name of the output parser

**Supports Streaming**: Whether the output parser supports streaming.

**Input Type**: Expected input type. Most output parsers work on both strings and messages, but some (like OpenAI Functions) need a message with specific arguments.

**Output Type**: The output type of the object returned by the parser.

**Description**: Our commentary on this output parser and when to use it.

The current date is ${new Date().toISOString()}`;

// Noop statement to hide output
void 0;

const modelWithCaching = new ChatOpenAI({
  model: "gpt-5.4-mini",
});

// CACHED_TEXT is some string longer than 1024 tokens
const LONG_TEXT = `You are a pirate. Always respond in pirate dialect.

Use the following as context when answering questions:

${CACHED_TEXT}`;

const longMessages = [
  {
    role: "system",
    content: LONG_TEXT,
  },
  {
    role: "user",
    content: "What types of messages are supported in LangChain?",
  },
];

const originalRes = await modelWithCaching.invoke(longMessages);

console.log("USAGE:", originalRes.response_metadata.usage);

USAGE: {
  prompt_tokens: 2624,
  completion_tokens: 263,
  total_tokens: 2887,
  prompt_tokens_details: { cached_tokens: 0 },
  completion_tokens_details: { reasoning_tokens: 0 }
}

const resWitCaching = await modelWithCaching.invoke(longMessages);

console.log("USAGE:", resWitCaching.response_metadata.usage);

USAGE: {
  prompt_tokens: 2624,
  completion_tokens: 272,
  total_tokens: 2896,
  prompt_tokens_details: { cached_tokens: 2432 },
  completion_tokens_details: { reasoning_tokens: 0 }
}

预测输出

某些 OpenAI 模型（如其 gpt-4o 和 gpt-4o-mini 系列）支持预测输出，允许您提前传入 LLM 预期输出的已知部分以减少延迟。这对于编辑文本或代码等场景非常有用，因为模型输出中只有一小部分会发生变化。

示例如下：

const modelWithPredictions = new ChatOpenAI({
  model: "gpt-5.4-mini",
});

const codeSample = `
/// <summary>
/// Represents a user with a first name, last name, and username.
/// </summary>
public class User
{
/// <summary>
/// Gets or sets the user's first name.
/// </summary>
public string FirstName { get; set; }

/// <summary>
/// Gets or sets the user's last name.
/// </summary>
public string LastName { get; set; }

/// <summary>
/// Gets or sets the user's username.
/// </summary>
public string Username { get; set; }
}
`;

// Can also be attached ahead of time
// using `model.bind({ prediction: {...} })`;
await modelWithPredictions.invoke(
  [
    {
      role: "user",
      content:
        "Replace the Username property with an Email property. Respond only with code, and with no markdown formatting.",
    },
    {
      role: "user",
      content: codeSample,
    },
  ],
  {
    prediction: {
      type: "content",
      content: codeSample,
    },
  }
);

AIMessage {
  "id": "chatcmpl-AQLyQKnazr7lEV7ejLTo1UqhzHDBl",
  "content": "/// <summary>\n/// Represents a user with a first name, last name, and email.\n/// </summary>\npublic class User\n{\n/// <summary>\n/// Gets or sets the user's first name.\n/// </summary>\npublic string FirstName { get; set; }\n\n/// <summary>\n/// Gets or sets the user's last name.\n/// </summary>\npublic string LastName { get; set; }\n\n/// <summary>\n/// Gets or sets the user's email.\n/// </summary>\npublic string Email { get; set; }\n}",
  "additional_kwargs": {},
  "response_metadata": {
    "tokenUsage": {
      "promptTokens": 148,
      "completionTokens": 217,
      "totalTokens": 365
    },
    "finish_reason": "stop",
    "usage": {
      "prompt_tokens": 148,
      "completion_tokens": 217,
      "total_tokens": 365,
      "prompt_tokens_details": {
        "cached_tokens": 0
      },
      "completion_tokens_details": {
        "reasoning_tokens": 0,
        "accepted_prediction_tokens": 36,
        "rejected_prediction_tokens": 116
      }
    },
    "system_fingerprint": "fp_0ba0d124f1"
  },
  "tool_calls": [],
  "invalid_tool_calls": [],
  "usage_metadata": {
    "output_tokens": 217,
    "input_tokens": 148,
    "total_tokens": 365,
    "input_token_details": {
      "cache_read": 0
    },
    "output_token_details": {
      "reasoning": 0
    }
  }
}

请注意，当前预测会按额外 token 计费，虽然减少了延迟，但会增加您的用量和成本。

音频输出

某些 OpenAI 模型（如 gpt-4o-audio-preview）支持生成音频输出。此示例展示如何使用该功能：

const modelWithAudioOutput = new ChatOpenAI({
  model: "gpt-4o-audio-preview",
  // You may also pass these fields to `.bind` as a call argument.
  modalities: ["text", "audio"], // Specifies that the model should output audio.
  audio: {
    voice: "alloy",
    format: "wav",
  },
});

const audioOutputResult = await modelWithAudioOutput.invoke("Tell me a joke about cats.");
const castAudioContent = audioOutputResult.additional_kwargs.audio as Record<string, any>;

console.log({
  ...castAudioContent,
  data: castAudioContent.data.slice(0, 100) // Sliced for brevity
})

{
  id: 'audio_67129e9466f48190be70372922464162',
  data: 'UklGRgZ4BABXQVZFZm10IBAAAAABAAEAwF0AAIC7AAACABAATElTVBoAAABJTkZPSVNGVA4AAABMYXZmNTguMjkuMTAwAGRhdGHA',
  expires_at: 1729277092,
  transcript: "Why did the cat sit on the computer's keyboard? Because it wanted to keep an eye on the mouse!"
}

我们可以看到音频数据返回在 data 字段中。我们还获得了一个 expires_at 日期字段。此字段表示音频响应在服务器上不再可访问以用于多轮对话的日期。

流式音频输出

OpenAI也支持流式音频输出。以下是一个示例：

const modelWithStreamingAudioOutput = new ChatOpenAI({
  model: "gpt-4o-audio-preview",
  modalities: ["text", "audio"],
  audio: {
    voice: "alloy",
    format: "pcm16", // Format must be `pcm16` for streaming
  },
});

const audioOutputStream = await modelWithStreamingAudioOutput.stream("Tell me a joke about cats.");
let finalAudioOutputMsg: AIMessageChunk | undefined;
for await (const chunk of audioOutputStream) {
  finalAudioOutputMsg = finalAudioOutputMsg ? concat(finalAudioOutputMsg, chunk) : chunk;
}
const castStreamedAudioContent = finalAudioOutputMsg?.additional_kwargs.audio as Record<string, any>;

console.log({
  ...castStreamedAudioContent,
  data: castStreamedAudioContent.data.slice(0, 100) // Sliced for brevity
})

{
  id: 'audio_67129e976ce081908103ba4947399a3eaudio_67129e976ce081908103ba4947399a3e',
  transcript: 'Why was the cat sitting on the computer? Because it wanted to keep an eye on the mouse!',
  index: 0,
  data: 'CgAGAAIADAAAAA0AAwAJAAcACQAJAAQABQABAAgABQAPAAAACAADAAUAAwD8/wUA+f8MAPv/CAD7/wUA///8/wUA/f8DAPj/AgD6',
  expires_at: 1729277096
}

音频输入

这些模型还支持传递音频作为输入。为此，您必须指定 input_audio 字段，如下所示：

const userInput = new HumanMessage({
  content: [{
    type: "input_audio",
    input_audio: {
      data: castAudioContent.data, // Reuse the base64 data from the first example
      format: "wav",
    },
  }]
})

// Reuse the same model instance
const userInputAudioRes = await modelWithAudioOutput.invoke([userInput]);

console.log((userInputAudioRes.additional_kwargs.audio as Record<string, any>).transcript);

That's a great joke! It's always fun to imagine why cats do the funny things they do. Keeping an eye on the "mouse" is a creatively punny way to describe it!

API 参考

有关所有 ChatOpenAI 功能和配置的详细文档，请前往 API 参考.