跳至主要内容

如何获取日志概率

先决条件

本指南假设您熟悉以下概念

某些聊天模型可以配置为返回令牌级日志概率，表示给定令牌的可能性。本指南将介绍如何在 LangChain 中获取此信息。

OpenAI​

安装 @langchain/openai 包并设置您的 API 密钥

提示

请参阅此部分了解有关安装集成包的一般说明。

yarn add @langchain/openai @langchain/core

为了使 OpenAI API 返回日志概率，我们需要将 logprobs 参数设置为 true。然后，日志概率将包含在每个输出 AIMessage 中，作为 response_metadata 的一部分

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

const model = new ChatOpenAI({
  model: "gpt-4o",
  logprobs: true,
});

const responseMessage = await model.invoke("how are you today?");

responseMessage.response_metadata.logprobs.content.slice(0, 5);
[
  {
    token: "Thank",
    logprob: -0.70174205,
    bytes: [ 84, 104, 97, 110, 107 ],
    top_logprobs: []
  },
  {
    token: " you",
    logprob: 0,
    bytes: [ 32, 121, 111, 117 ],
    top_logprobs: []
  },
  {
    token: " for",
    logprob: -0.000004723352,
    bytes: [ 32, 102, 111, 114 ],
    top_logprobs: []
  },
  {
    token: " asking",
    logprob: -0.0000013856493,
    bytes: [
       32,  97, 115,
      107, 105, 110,
      103
    ],
    top_logprobs: []
  },
  {
    token: "!",
    logprob: -0.00030102333,
    bytes: [ 33 ],
    top_logprobs: []
  }
]

并且也作为流式传输的消息块的一部分

let count = 0;
const stream = await model.stream("How are you today?");
let aggregateResponse;

for await (const chunk of stream) {
  if (count > 5) {
    break;
  }
  if (aggregateResponse === undefined) {
    aggregateResponse = chunk;
  } else {
    aggregateResponse = aggregateResponse.concat(chunk);
  }
  console.log(aggregateResponse.response_metadata.logprobs?.content);
  count++;
}
[]
[
  {
    token: "Thank",
    logprob: -0.23375113,
    bytes: [ 84, 104, 97, 110, 107 ],
    top_logprobs: []
  }
]
[
  {
    token: "Thank",
    logprob: -0.23375113,
    bytes: [ 84, 104, 97, 110, 107 ],
    top_logprobs: []
  },
  {
    token: " you",
    logprob: 0,
    bytes: [ 32, 121, 111, 117 ],
    top_logprobs: []
  }
]
[
  {
    token: "Thank",
    logprob: -0.23375113,
    bytes: [ 84, 104, 97, 110, 107 ],
    top_logprobs: []
  },
  {
    token: " you",
    logprob: 0,
    bytes: [ 32, 121, 111, 117 ],
    top_logprobs: []
  },
  {
    token: " for",
    logprob: -0.000004723352,
    bytes: [ 32, 102, 111, 114 ],
    top_logprobs: []
  }
]
[
  {
    token: "Thank",
    logprob: -0.23375113,
    bytes: [ 84, 104, 97, 110, 107 ],
    top_logprobs: []
  },
  {
    token: " you",
    logprob: 0,
    bytes: [ 32, 121, 111, 117 ],
    top_logprobs: []
  },
  {
    token: " for",
    logprob: -0.000004723352,
    bytes: [ 32, 102, 111, 114 ],
    top_logprobs: []
  },
  {
    token: " asking",
    logprob: -0.0000029352968,
    bytes: [
       32,  97, 115,
      107, 105, 110,
      103
    ],
    top_logprobs: []
  }
]
[
  {
    token: "Thank",
    logprob: -0.23375113,
    bytes: [ 84, 104, 97, 110, 107 ],
    top_logprobs: []
  },
  {
    token: " you",
    logprob: 0,
    bytes: [ 32, 121, 111, 117 ],
    top_logprobs: []
  },
  {
    token: " for",
    logprob: -0.000004723352,
    bytes: [ 32, 102, 111, 114 ],
    top_logprobs: []
  },
  {
    token: " asking",
    logprob: -0.0000029352968,
    bytes: [
       32,  97, 115,
      107, 105, 110,
      103
    ],
    top_logprobs: []
  },
  {
    token: "!",
    logprob: -0.00039694557,
    bytes: [ 33 ],
    top_logprobs: []
  }
]

topLogprobs​

要查看每个步骤的替代潜在生成，可以使用 topLogprobs 参数

const modelWithTopLogprobs = new ChatOpenAI({
  model: "gpt-4o",
  logprobs: true,
  topLogprobs: 3,
});

const res = await modelWithTopLogprobs.invoke("how are you today?");

res.response_metadata.logprobs.content.slice(0, 5);
[
  {
    token: "I'm",
    logprob: -2.2864406,
    bytes: [ 73, 39, 109 ],
    top_logprobs: [
      {
        token: "Thank",
        logprob: -0.28644064,
        bytes: [ 84, 104, 97, 110, 107 ]
      },
      {
        token: "Hello",
        logprob: -2.0364406,
        bytes: [ 72, 101, 108, 108, 111 ]
      },
      { token: "I'm", logprob: -2.2864406, bytes: [ 73, 39, 109 ] }
    ]
  },
  {
    token: " just",
    logprob: -0.14442946,
    bytes: [ 32, 106, 117, 115, 116 ],
    top_logprobs: [
      {
        token: " just",
        logprob: -0.14442946,
        bytes: [ 32, 106, 117, 115, 116 ]
      },
      { token: " an", logprob: -2.2694294, bytes: [ 32, 97, 110 ] },
      {
        token: " here",
        logprob: -4.0194297,
        bytes: [ 32, 104, 101, 114, 101 ]
      }
    ]
  },
  {
    token: " a",
    logprob: -0.00066632946,
    bytes: [ 32, 97 ],
    top_logprobs: [
      { token: " a", logprob: -0.00066632946, bytes: [ 32, 97 ] },
      {
        token: " lines",
        logprob: -7.750666,
        bytes: [ 32, 108, 105, 110, 101, 115 ]
      },
      { token: " an", logprob: -9.250667, bytes: [ 32, 97, 110 ] }
    ]
  },
  {
    token: " computer",
    logprob: -0.015423919,
    bytes: [
       32,  99, 111, 109,
      112, 117, 116, 101,
      114
    ],
    top_logprobs: [
      {
        token: " computer",
        logprob: -0.015423919,
        bytes: [
           32,  99, 111, 109,
          112, 117, 116, 101,
          114
        ]
      },
      {
        token: " program",
        logprob: -5.265424,
        bytes: [
           32, 112, 114, 111,
          103, 114,  97, 109
        ]
      },
      {
        token: " machine",
        logprob: -5.390424,
        bytes: [
           32, 109,  97,  99,
          104, 105, 110, 101
        ]
      }
    ]
  },
  {
    token: " program",
    logprob: -0.0010724656,
    bytes: [
       32, 112, 114, 111,
      103, 114,  97, 109
    ],
    top_logprobs: [
      {
        token: " program",
        logprob: -0.0010724656,
        bytes: [
           32, 112, 114, 111,
          103, 114,  97, 109
        ]
      },
      {
        token: "-based",
        logprob: -6.8760724,
        bytes: [ 45, 98, 97, 115, 101, 100 ]
      },
      {
        token: " algorithm",
        logprob: -10.626073,
        bytes: [
           32,  97, 108, 103,
          111, 114, 105, 116,
          104, 109
        ]
      }
    ]
  }
]

后续步骤​

现在您已经了解了如何在 LangChain 中从 OpenAI 模型获取日志概率。

接下来，查看本部分中的其他操作指南聊天模型，例如如何让模型返回结构化输出或如何跟踪令牌使用情况。

此页面是否有帮助？


您也可以留下详细的反馈 在 GitHub 上.