feat(translator): remove LLM <think>xxx</think>

[missyoueveryday]

问题描述

使用ollama本地模型进行翻译，如果使用带有思考的模型比如deepseek r1会将思考部分也输入到翻译结果中，并且同时导致排版错误

可用于测试的pdf：

test.pdf

测试文档

Important

请提供用于复现测试的 PDF 文档

[awwaawwa]

There is currently an issue, which is that if the original text actually contains <think>xxx</think>, it will be mistakenly affected by the regex.

[missyoueveryday]

easy way is after get the response, add a filter to delete the 'think' part and then go through the normal output.
Is this a good solution?

[awwaawwa]

The current issue is how to distinguish between the <think>xxx</think> generated by the model and the <think>xxx</think> contained in the original text.

[awwaawwa]

If unable to distinguish this, it may accidentally remove the <think>xxx</think> in the original text

[awwaawwa]

Using ^<think>.+</think> should be able to reduce the probability of false positives to a very low level. Looking forward to someone who can implement this.

[missyoueveryday]

How about simply adding a checkbox to let users choose whether the model being used is a deep thinking model? 😎
default be normal model

[awwaawwa]

Since the false positive rate of using ^<think>.+</think> should be low enough, there's no need to add a checkbox. Additionally, the current configuration system of pdf2zh is not very convenient for adding parameters.......

[awwaawwa]

#586 After rewriting, adding parameters becomes much more convenient.

[hellofinch]

ollama没有关闭think输出的开关，所以只能从试着快刀斩乱麻，粗暴的正则处理了。