I’m working on API project where I’d like to have ChatGPT ask my leads basic questions and once the basics are collected, I want ChatGPT to return something like “I_AM_DONE_WITH_COLLECTING_ALL_THE_BASIC_INFO” so I can transition the conversation to a real human being in our office.
I’d be listening for my CRM events with webhooks and I could catch the I_AM_DONE_WITH_COLLECTING_ALL_THE_BASIC_INFO string in lead’s notes/communications and forward the conversation to a real human being relatively easily.
More details about the project:
- A lead comes into my CRM
- A conversation starts via SMS/text messaging. On my side, it would be ChatGPT asking the questions. In the other side, it’s the lead = the customer.
- I need to know what kind of problem they have - getting the information about the problem ifself is not straightfoward but so far the question’s I’ve been getting (while playing a fake lead) in the ChatGPT’s web interface were prety good.
- Where they live
- What’s their email etc.
- The point I’m struggling the most is how to make GhatGPT “realize” it has collected all the information so I can transition the chat to a real human being (if the lead is good). And by “realizing”, I want ChatGPT to return the I_AM_DONE_WITH_COLLECTING_ALL_THE_BASIC_INFO string.
Here are my problems:
Problem 1:
I tried to run a few chats where I instructed ChatGPT to return specific string when all the info is collected (the I_AM_DONE_WITH_COLLECTING_ALL_THE_BASIC_INFO) and it would sometimes work but sometimes it would just continue the conversation with the lead despite already having all the info needed. I literally said “once you have all the information collected, return I_AM_DONE_WITH_COLLECTING_ALL_THE_BASIC_INFO and do not engage in the chat anymore”.
Problem 2:
I am also struggling with making ChatGPT to ask just one question at a time. Again, I would intruct “you are only allowed to ask one question at a time” but it would sometimes work but sometimes I’d get 7 questions at once.
I’m wondering whether anyone has been facing the same challenge and was able to find any reasonable solution to these problems?
And yes, I do realize maintaing a conversation through API is not that easy because I will need to send all the previous replies for every new prompt API request but let’s say it can be done.