How do you get clean results for rewriting?

[JonatanSahar]

I can't seem to make any of the LLMs I'm working with to give output which I can directly incorporate into my code as accepted diff.. at the minimum I get code fences, but almost always I get a bunch of text, even when my prompt is something like:

(refactor . "You are a language model with expert programming knowledge. Provide clean, correct code solutions with minimal commentary; output code and only code, DO NOT add code fences e.g. Python ''' ''' around the code.")

Thanks!

[karthink]

Is this with gptel-rewrite or regular use of gptel-send?

[JonatanSahar]

Both, as well as with the rewrite option through gptel-send.

[jdm204]

I'm having the same issue, always getting markdown code fences when I ask for Julia code to be rewritten, despite the instructions being:

You are a julia programmer.  Follow my instructions and refactor julia code I provide.
- Generate ONLY julia code as output, without any explanation or markdown code fences.
- Generate code in full, do not abbreviate or omit code.
- Do not ask for further clarification, and make any assumptions you need to follow instructions.

[karthink]

I haven't faced this issue, but I haven't tried the rewrite operation with the smaller/local open-weight models.

You can work around this via post-processing the response in gptel-post-rewrite-functions (for gptel-rewrite) and gptel-post-response-functions (for gptel-send). You just have to check for triple backticks in the first and last line of the response and remove them if they are present. If you don't want to write your own helper, there are various versions of this code spread across the Github issue tracker.