使用 deepseek r1 作为【话题命名模型】时，话题名称会出现深度思考的信息。

[gavinliu]

📦 部署环境

Other, Docker

📌 软件版本

1.51.1

💻 系统环境

macOS

🌐 浏览器

Edge

🐛 问题描述

使用 deepseek r1 作为【话题命名模型】时，话题名称会出现深度思考的信息。

📷 复现步骤

1. 话题命名模型设置成 deepseek r1
2. 问答完成后，观察话题命名

如下：

<think> 好的，我需要将用户的会话内容总结为一个不超过10个字的中文标题。首先，用户询问如何用C++编写跨平台读取图片EXIF信息的程序。这涉及到使用特定库、处理文件路径、注释代码以及错误处理等方面。 接下来，我查看了回答部分，发现提到了使用libexif库，并详细说明了项目结构、依赖项、代码实现和测试等。这些内容都围绕如何编写一个支持跨平台且读取EXIF信息的C++程序展开。 为了生成标题，需要抓住关键点：C++编程、图片处理、EXIF信息、跨平台。考虑到字数限制，选择最核心的部分来表达。最终决定使用“用C++编写跨平台读取EXIF信息程序”作为标题，它简洁明了地概括了整个主题。 </think> 用C++编写跨平台读取EXIF信息程序

🚦 期望结果

期望排除深度思考标签，直接显示结果。

如下：

用C++编写跨平台读取EXIF信息程序

📝 补充信息

No response

[lobehubbot]

👀 @gavinliu

Thank you for raising an issue. We will investigate into the matter and get back to you as soon as possible.
Please make sure you have given us as much context as possible.
非常感谢您提交 issue。我们会尽快调查此事，并尽快回复您。 请确保您已经提供了尽可能多的背景信息。

[lobehubbot]

Bot detected the issue body's language is not English, translate it automatically. 👯👭🏻🧑‍🤝‍🧑👫🧑🏿‍🤝‍🧑🏻👩🏾‍🤝‍👨🏿👬🏿

---

📦 deployment environment

Other, Docker

📌 Software version

1.51.1

💻 System environment

macos

🌐 browser

EDGE

🐛 Problem description

When using deepseek r1 as the [Topic Naming Model], deep thinking information will appear on the topic name.

📷 Reapped steps

1. Set the topic naming model to deepseek r1
2. After the Q&A is completed, observe the topic name

as follows:

<think> OK, I need to summarize the user's session content into a Chinese title with no more than 10 words. First, the user asks how to write a program that uses C++ to read image EXIF ​​information across platforms. This involves the use of specific libraries, processing file paths, commenting code, and error handling. Next, I looked at the answer section and found that using the libexif library mentioned, and detailed the project structure, dependencies, code implementation, and testing, etc. All of these revolves around how to write a C++ program that supports cross-platform and reads EXIF ​​information. In order to generate titles, we need to grasp the key points: C++ programming, picture processing, EXIF ​​information, and cross-platform. Considering the word count limitation, select the most core part to express. Finally, I decided to use the title "Writing a Cross-Platform Read EXIF ​​Information Program in C++" and it summarizes the entire topic concisely and clearly. </think> Write a cross-platform read EXIF ​​information program in C++

🚦 expectation result

Expect to exclude deep thinking labels and directly display the results.

as follows:

Write a cross-platform EXIF ​​information program in C++

📝 Supplement information

No response

[dosubot[bot]]

@gavinliu Hello! I'm here to help you with your software issue. If you need assistance with bugs, questions, or contributions, just let me know!

The issue you're experiencing is related to the inclusion of <think> tags in the output when using the DeepSeek R1 model for topic naming. This behavior is specific to the model and cannot be prevented through Lobe Chat's application layer ^[1].

To address this, you can use the normalizeThinkTags function, which is designed to handle <think> tags by ensuring they have two line breaks before and after them, and it also removes any potential extra line breaks. You can modify this function to completely remove the <think> tags if that's your desired outcome. Here is the function as it currently exists:

export const normalizeThinkTags = (input: string) => {
  return (
    input
      .replaceAll(/([^\n])\s*<think>/g, '$1\n\n<think>')
      .replaceAll(/<think>\s*([^\n])/g, '<think>\n\n$1')
      .replaceAll(/([^\n])\s*<\/think>/g, '$1\n\n</think>')
      .replaceAll(/<\/think>\s*([^\n])/g, '</think>\n\n$1')
      .replaceAll(/\n{3,}/g, '\n\n')
  );
};

You can adapt this function to remove the <think> tags entirely by replacing them with an empty string.

To continue talking to Dosu, mention @dosu.

---

^{Help Dosu learn! Give it feedback: Great Response | Irrelevant Answer | Incorrect Sources | Too Verbose | Hallucination | Other}

[arvinxx]

如果你用 DS 官方或者硅基流动的 R1 模型，不会出现这个问题。你用了例如 ollama 部署的模型才会有这个问题。

暂时不考虑移除这类 think 标签的内容，建议不要用 Ollama 等部署的深度思考的模型作为话题命名模型。