I am trying to make a Poe bot that generates an email based on the sender, recipient, and subject of the email. Despite including the lines “Do not include any information aside from what is explicitly given to you by the user. Do not assume anything. If you feel that more information about something would be particularly helpful in writing the email, please request additional information from the user” in the prompt, the bot often assumes things when writing the emails. For example, when trying to generate an email asking for time off, it will usually assume that the writer has spoken with coworkers about covering shifts/missed work. Does anyone have a good way to prevent this from happening? I thought those lines would be explicit enough that it would stop the problem.
I feel like this is something you should post on the Poe forums.
Ignoring Poe, what’s your temperature? Why not run it through a second prompt which catches any assumed context, and then refines the email?
But, if you just want a presumption-free response you could just have the bot select a template and fill it in. It sounds like you just want some uniquely wordspun content.
As a general rule of thumb, GPT models do not listen to do not. But they do take guidance very well.
I could not find any Poe forums, and Poe doesn’t provide a way to control the temperature. The goal of the bot is to take in a message and generate an email, asking for more information if it feels it’s necessary. The purpose is to assist in the nuances of tone for those who may not be familiar with the proper procedures, such as people whose first language is not English or non-neurotypical people. Running it through a second prompt is not an option on Poe.
As for not listening to do not, that is interesting. I figured the sentence “If you feel that more information about something would be particularly helpful in writing the email, please request additional information from the user” was that method of guidance and that having both the “do” and “do not” would be effective.
Those are some brutal restrictions.
It may be beneficial to try the OpenAI Playground. There’s a little more control and clarity for debugging.
This is very subjective and open for interpretation. Have you tried it without this part included?
Has it ever asked for more information?
You should for sure try using different words. Have you tried for example saying “Use only those informations that are given to you. Never use informations that you don’t know of what means you should not assume anything as well. If there are needed any other informations in blank space use [ ] and ask for what you seems that should be written there as well”. You should know that To make poe understand you just fine you should think of many aspects. Now that you know that it is assuming that the writer has spoken with coworkers about covering shifts/missed work, you can tell it not to do this specific thing. Remember to use the most important informations at the top and go so on with less important informations. This thing should work in my opinion.