Hi, I’m running into a bizarre issue which is driving me crazy. I have a custom GPT using OAuth to connect to our app. It works pretty well, right up until I upload a Knowledge file. I already have given it an openapi schema, which it uses pretty effectively, but I wanted to give it more context around example commands a user might give, and the api endpoints it should use, and the parameters it should send. After uploading, it would no longer make successful POST calls to one of our API endpoints. I thought maybe somehow the content of the file was overriding or conflicting with other instructions that made it work before, so I dumbed down the file to basically not give it any meaningful instructions (it wouldn’t let me upload an empty file). Even with essentially nothing in the file, the mere presence of the file screws up the api request. I’ve tested removing it and adding it back multiple times, and it consistently works without it, and not with it. GPT says it is sending the request, and when I click on debug it seems to show the appropriate parameters. But it does not show a response received, and our backend logs don’t show any request being made (even if the request was misformatted we would still see it there). Any thoughts on why this is happening and how to fix it? Should I just cram everything into the instructions field instead? (seems like that could get a little out of hand as we expand the GPT’s capabilities)
1 Like
Knowledge is scary. The mere enablement of “code execution” starts your custom llm “executing” your knowledge, and then it starts writing bad python behind your back because it thinks it “knows” the path from your reference file and not your API. So then it starts giving you python-like gibberish and the little > prompt.
I can give it clear directives (ammusingly in python) in my configuration, but the longer my config gets, the less reliable the VM comes.
I use a file to pass along knowledge about how to use paramaters in my reduced subset API, and turning on code always borks me.