I have an issue with generating answers from a context. I am using assistants and I have setup a way to change an assistant’s instructions and add a bunch of files in its vector store for file search. These files contain urls - each file has at least one url. I have setup instructions so that in every response which searched files it needs to give the url from the file content. When answering it finds the correct files, however when providing the urls it very often slightly changes them as if treating the urls like common language. This results in broken url links which are unusable to me. I have tried to tell it in instructions to never modify urls but it just doesn’t listen and constantly changes them. I have tried models gpt-4o-mini and gpt-4o and both give the same result.
Can anyone help with this? How can I ensure that the model doesn’t modify the urls but just copy and pastes them from its knowledge?