I have created an assistant and have set instructions that it must not answer any questions beyond the context of the thread. I have asked it to assume the role of an evaluator of roman architecture, and the person on other side is a student. Now, how do I ask it a question like “Chatgpt, how do you like the student?”, how do I go under the facade of the evaluator and ask it questions that are direct and not from a student.
Hi!
As a straightforward but suboptimal solution you can add a secret code to the instructions that allows the assistant(?) to change it’s behavior when prompted by the user.
We can do better but you need to share some extra information about how you implemented the assistant/custom GPT. You are either referring to a custom GPT (ChatGPT interface) or an Assistant via the API.
Then, if you can share how you expect the model to know about it’s previous conversations with the students this will be most helpful.
Basically I am using the chat GPT API for assistants. I have several candidates that are being assessed by the same assistant in different threads. Instructions provided to the assistant can be summarised as, “assume you are an evaluator of roman architecture and do not answer anything related to the ongoing evaluation that may be asked by the candidate. You need to assess him on the scale of 1-10, each question he correctly or incorrectly answers can increment his current scale by 1 or decrement by one. If he stabilises at a point, he is done and you may thank him”
Now since this is time bound, and maybe 100 threads are active at any time against the same assistant, I have a master thread that periodically goes to each of these threads and enquires about the current scale of the candidate, if I find 5 such that are beyond level 7, everyone exits. Intent is to quickly determine the best 5 as soon as possible from a set of 100 candidates taking the test
1 Like
Ok. What about getting the messages from the student threads and passing these to an evaluator instance of the model?
This way you keep the student experience completely separated from the evaluation process.
Thank you for the reply, and I get what you mean, but is that not kinda brute force? And might deplete my token balance also
I used the assistant model because that decouples me from retaining message history. Here you are asking me to create a separate assistant model that is trained to “reveal” evaluation and send in all messages on the original thread to this one. Too much network back and forth also. I think there must be some simpler way to ask "@chatgpt this is me and tell me how this candidate is " on the original model/assistant. Is there?
If you are looking for a no-code solution you can fall back to the very first suggestion I made above and add a secret password to the instructions that will allow you to override the behavior aimed at the students.
In fact any technique to override the system prompt will do.
1 Like
Yes, the secret password works. Than you for the help
1 Like
This topic was automatically closed 2 days after the last reply. New replies are no longer allowed.