Perplexica is attempting to submit enormous blocks of AI thoughts to search

[nomonkeynodeal]

It seems to be working well for me, but then I found this in my Searxng logs. My LLM is ExLlamaV3 with GLM 4.5 air. This is a thinking model. No errors in the logs on the LLM side. I'm not sure what other details I should add, but please ask.

[ItzCrazyKns]

The problem is: Langchain doesn't provide a method or a global way to separate reasoning from the response. I'll look into what I can do.

[nomonkeynodeal]

I just encountered a similar error where the thinking was being randomly jammed into the searches, but this time it was more obvious because I was running llama.cpp/gpt-oss without --jinja, so the chat format was completely messed up.

I wonder if a more subtle problem with ExLlamaV3 think tags caused the above.

[ItzCrazyKns]

If think XML tag is returned, we can remove it but with some chat templates, it outputs it with the response and with in Langchain there's no standard for reasoning.

[ItzCrazyKns]

Related to langchain-ai/langchainjs#9180

[dogzilla]

Would seeing doubled responses from gpt-oss be the same issue?

[ItzCrazyKns]

Would seeing doubled responses from gpt-oss be the same issue?

The model might as well return what its about to answer in the reasoning content as well but we cannot say without looking at the text.