Hi
I’m uploading a MS Word document for a messaging framework for an Assistant to analyze using File search but am getting inconsistent results.
I reviewed a thread and after the file loads this line appears:
code_interpreter(msearch([“messaging framework”])
I’m not sure why the code_interpreter is being run for a file search - does any one know why and if I could/should be able to prevent this happening?
Thanks
Paul
This would appear to be the AI being very confused by the tools it has been provided.
msearch(queries: list[str]) is a method for myfiles_browser.
The comprehension and perplexity of how to do just about anything has been pulled way back on the latest models. You can set top_p: 0.3 so that a unlikely “code” token isn’t the first emitted after invoking a tool generation path.
You can imagine the AI at a following token, “what am I going to do now since I should search but the tool name to complete on is code interpreter…”
You can try gpt-4-1106-preview as long as you aren’t needing non-English accented characters written to tools.
Thanks for the response. I’ll try the preview version.