跳至主要内容

如何将工具输出传递给聊天模型

先决条件

本指南假设您熟悉以下概念

某些模型能够进行 工具调用 - 生成符合特定用户提供的模式的参数。本指南将演示如何使用这些工具调用来实际调用函数，并将结果正确地传递回模型。

首先，让我们定义我们的工具和模型

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

const addTool = tool(
  async ({ a, b }) => {
    return a + b;
  },
  {
    name: "add",
    schema: z.object({
      a: z.number(),
      b: z.number(),
    }),
    description: "Adds a and b.",
  }
);

const multiplyTool = tool(
  async ({ a, b }) => {
    return a * b;
  },
  {
    name: "multiply",
    schema: z.object({
      a: z.number(),
      b: z.number(),
    }),
    description: "Multiplies a and b.",
  }
);

const tools = [addTool, multiplyTool];

选择您的聊天模型

安装依赖项

提示

参见 本节介绍有关安装集成包的常规说明.

yarn add @langchain/openai

添加环境变量

OPENAI_API_KEY=your-api-key

实例化模型

import { ChatOpenAI } from "@langchain/openai";

const llm = new ChatOpenAI({
  model: "gpt-4o-mini",
  temperature: 0
});

现在，让我们让模型调用一个工具。我们将其添加到一个消息列表中，我们将将其视为对话历史记录

import { HumanMessage } from "@langchain/core/messages";

const llmWithTools = llm.bindTools(tools);

const messages = [new HumanMessage("What is 3 * 12? Also, what is 11 + 49?")];

const aiMessage = await llmWithTools.invoke(messages);

console.log(aiMessage);

messages.push(aiMessage);
AIMessage {
  "id": "chatcmpl-9p1NbC7sfZP0FE0bNfFiVYbPuWivg",
  "content": "",
  "additional_kwargs": {
    "tool_calls": [
      {
        "id": "call_RbUuLMYf3vgcdSQ8bhy1D5Ty",
        "type": "function",
        "function": "[Object]"
      },
      {
        "id": "call_Bzz1qgQjTlQIHMcEaDAdoH8X",
        "type": "function",
        "function": "[Object]"
      }
    ]
  },
  "response_metadata": {
    "tokenUsage": {
      "completionTokens": 50,
      "promptTokens": 87,
      "totalTokens": 137
    },
    "finish_reason": "tool_calls",
    "system_fingerprint": "fp_400f27fa1f"
  },
  "tool_calls": [
    {
      "name": "multiply",
      "args": {
        "a": 3,
        "b": 12
      },
      "type": "tool_call",
      "id": "call_RbUuLMYf3vgcdSQ8bhy1D5Ty"
    },
    {
      "name": "add",
      "args": {
        "a": 11,
        "b": 49
      },
      "type": "tool_call",
      "id": "call_Bzz1qgQjTlQIHMcEaDAdoH8X"
    }
  ],
  "invalid_tool_calls": [],
  "usage_metadata": {
    "input_tokens": 87,
    "output_tokens": 50,
    "total_tokens": 137
  }
}
2

接下来，让我们使用模型填充的参数来调用工具函数！

方便的是，如果我们用 ToolCall 调用 LangChain Tool，我们将自动获得一个可以反馈给模型的 ToolMessage

兼容性

此功能需要 @langchain/core>=0.2.16。有关升级的指南，请参见此处 。

如果您使用的是 @langchain/core 的早期版本，则需要手动使用工具调用的字段来构建 ToolMessage。

const toolsByName = {
  add: addTool,
  multiply: multiplyTool,
};

for (const toolCall of aiMessage.tool_calls) {
  const selectedTool = toolsByName[toolCall.name];
  const toolMessage = await selectedTool.invoke(toolCall);
  messages.push(toolMessage);
}

console.log(messages);
[
  HumanMessage {
    "content": "What is 3 * 12? Also, what is 11 + 49?",
    "additional_kwargs": {},
    "response_metadata": {}
  },
  AIMessage {
    "id": "chatcmpl-9p1NbC7sfZP0FE0bNfFiVYbPuWivg",
    "content": "",
    "additional_kwargs": {
      "tool_calls": [
        {
          "id": "call_RbUuLMYf3vgcdSQ8bhy1D5Ty",
          "type": "function",
          "function": "[Object]"
        },
        {
          "id": "call_Bzz1qgQjTlQIHMcEaDAdoH8X",
          "type": "function",
          "function": "[Object]"
        }
      ]
    },
    "response_metadata": {
      "tokenUsage": {
        "completionTokens": 50,
        "promptTokens": 87,
        "totalTokens": 137
      },
      "finish_reason": "tool_calls",
      "system_fingerprint": "fp_400f27fa1f"
    },
    "tool_calls": [
      {
        "name": "multiply",
        "args": {
          "a": 3,
          "b": 12
        },
        "type": "tool_call",
        "id": "call_RbUuLMYf3vgcdSQ8bhy1D5Ty"
      },
      {
        "name": "add",
        "args": {
          "a": 11,
          "b": 49
        },
        "type": "tool_call",
        "id": "call_Bzz1qgQjTlQIHMcEaDAdoH8X"
      }
    ],
    "invalid_tool_calls": [],
    "usage_metadata": {
      "input_tokens": 87,
      "output_tokens": 50,
      "total_tokens": 137
    }
  },
  ToolMessage {
    "content": "36",
    "name": "multiply",
    "additional_kwargs": {},
    "response_metadata": {},
    "tool_call_id": "call_RbUuLMYf3vgcdSQ8bhy1D5Ty"
  },
  ToolMessage {
    "content": "60",
    "name": "add",
    "additional_kwargs": {},
    "response_metadata": {},
    "tool_call_id": "call_Bzz1qgQjTlQIHMcEaDAdoH8X"
  }
]

最后，我们将使用工具结果调用模型。模型将使用此信息来生成我们原始查询的最终答案

await llmWithTools.invoke(messages);
AIMessage {
  "id": "chatcmpl-9p1NttGpWjx1cQoVIDlMhumYq12Pe",
  "content": "3 * 12 is 36, and 11 + 49 is 60.",
  "additional_kwargs": {},
  "response_metadata": {
    "tokenUsage": {
      "completionTokens": 19,
      "promptTokens": 153,
      "totalTokens": 172
    },
    "finish_reason": "stop",
    "system_fingerprint": "fp_18cc0f1fa0"
  },
  "tool_calls": [],
  "invalid_tool_calls": [],
  "usage_metadata": {
    "input_tokens": 153,
    "output_tokens": 19,
    "total_tokens": 172
  }
}

请注意，每个 ToolMessage 必须包含一个与模型生成的原始工具调用中的 id 匹配的 tool_call_id。这有助于模型将工具响应与工具调用匹配。

工具调用代理（如 LangGraph 中的代理）使用此基本流程来回答查询并解决任务。

您现在已经了解了如何将工具调用传递回模型。

以下指南可能会让您感兴趣

此页面是否有帮助？


您也可以留下详细的反馈 在 GitHub 上.