I have an assistant with the retrieval tool and some files associated with it. I want the assistant to use the information in the files for RAG but this should be done behind the scenes. What I see is that the assistant often says things like: Hello! I see that you've uploaded several files. How can I assist you today with these documents?
Has someone found a way to prevent the assistant from saying things like that? In our use case, the assistant and files are created beforehand and not uploaded by our users. I have tried adding things like this to the instructions but had no luck: The user did not upload any files, therefore you must never mention them. They should simply be used by you, behind the scenes, to obtain useful information relevant to the conversation. Never make reference to any file directly.
Give it a try with several models - they all act different.
And make the instructions very specific. Maybe like âYou are an assistant that has access to files that help you answer user questions. You should use the information in these files to come to your answers but do not ever add annotations, sources of any type and never mention of these files. As far as the user is concerned you just have all this knowledge.â
Or simply more information âfiles have been added to conversation history by the programmer, not your user. Thus, do not mention âfilesâ; only use the contents to enhance knowledgeâ
Thanks @jlvanhulst . I tried simply changing the instructions but it wasnât enough to change this behaviour. Changing the model to gpt-4-1106-preview seems to solve the problem though itâs not something I can afford at this time.
Thanks for pointing in the right direction, this would be the solution if I could go this path.
Thanks for the tip. Unfortunately this by itself didnât solve the problem. I also had to change the model to gpt-4-1106-preview but itâs not something I can do.
Either that preview model or gpt-3.5-turbo-1106 are the only ones that can understand and emit parallel tool function calls by their new training, which I suspect is required by some undisclosed function methods used in assistants.
addâl: you have to look deeper to endpoint compatibility to find the correct mention: 1106 models are required for âretrievalâ, in assistants, so this somewhat confirms the use of parallel calls (not primarily embedding) with undocumented methods within, only capable by the newest models.
in reality Assistant is a front end / consumer tool and thus your control is limited, as are your logs, your update path., your overall Dev., your nuanced prompts.
Imho⌠I would suggest that you scrap Assistant Dev and move to the full API since like a bad guitar you will never be able to take an in-tune tune out of Assistant. This is purely from my own perspective regarding intermediate / transitory tools / functionality.