Please see this post:
On 02/27/2024, ChatGPT 4 refused to read selected sections of public domain laws verbatim, even upon my request for discrete sections of text.
Another issue is with using open-source documentation - even when working around the verbatim constraint, I quite literally had to walk ChatGPT-4 step-by-step through the process of explaining changes to tabs between KivyMD v1.2.0 and v2.0.1 (meaning I pretty much had to complete the task myself, to then show ChatGPT how to do so - defeating the purpose of the exercise).
I had to explain that this didn’t require repeating any of the documentation verbatim - it just needed to base the response on the changes found in the documentation, and provide a code example.
Even specifying exactly which sections of the documentation/changelog to refer to did little to improve the response.
I did notice, however, that delimiting links like so <link.goes.here>, seemed to improve the response vs no delimiters.
I mean, a simple Google search returned better results with much less effort (without visiting the documentation itself). Not to mention telling me to refer to the official documentation every damn response, even when instructed not to do so.
Talk about frustrating.
We have been struggling with this.
A functional approach is to fill the input text with tags (i.e. ID-0, ID-1 after each dot or punctuation sign)
Then you can ask gpt to return the IDs of the selected text instead of a copy of it and it works much better
hope this helps