Files that are uploaded are either used for knowledge retrieval (and injection) or for code interpreter. Seemingly without differentiation, so attaching a file to an assistant or adding a file to a thread puts it both as something that can be accessed by code interpreter, and also puts into the retrieval system where it is loaded into the conversation context and uses tokens.
So how does one get an error (for which no predefined user-friendly message was written? You don’t refer to files if they are attached to an assistant for retrieval. They are simply always injected, or if too big for the model, a vector database fills the AI context length with chunks.
So it seems the likely path where “asking” could cause an error is related to code interpreter: either the model is trying to invoke python but has no code interpreter function, or something goes wrong with an enabled code interpreter that is returning an error (which the AI should learn from and iterate on).
It could even be hallucinating some “files” function.
If you aren’t using code interpreter, I would add additional prompting:
- python sandbox function is disabled for this session. you cannot call python to execute code;
- You have no files accessible by the user on demand. Any references to files shall be ignored; just answer from knowledge found in prior messages.