<think></think>
Directly deleting is too hasty. Can we optimize the scope of the deletion effect to only remove the character messages in the chat history? I’ve reported this issue on Bilibili and Xiaohongshu, but it hasn’t been changed.
<think></think>
Directly deleting is too hasty. Can we optimize the scope of the deletion effect to only remove the character messages in the chat history? I’ve reported this issue on Bilibili and Xiaohongshu, but it hasn’t been changed.
This tag will cause the DeepSeek API to report an error directly and must be removed. However, you can use details + regex for customization.
I recommend using regular expressions to handle the chain of thought. Let the AI not generate the <think></think> tags. Instead, generate <思维链></思维链>, or any other characters to wrap the chain of thought. Then use regular expressions to replace them with <think></think>. Specifically as follows:
Find regular expression: <思维链>([\s\S]*?)</思维链>
Replacement string: <think>$1</think>

Then you also need to create the same one, but change the position to “API request”. This way, the summary model will not see the chain of thought in the current reply of the role.
