As far as I got, ChatGpt plugins are called by chatGPT itself, but I also have seen another pattern that seems to reach a similar goal with a different approach : it is here at the certain point of the talk :
I do not take the plugin approach since they are not available yet for open.ai API
I think it’s easier to explain with an example
- get user question
- send to chatgpt :
-system message : Try to identify the intent of the user and reply accordingly given the examples and pattern below
- user message ( the “template” part) : if a user asks Find booking XYZKX or if a user asks Open booking XYZKX or or if a user asksI want to see booking XYZKX then replyWith : “{ "Intent" : "FindBooking" , "BookingId" : "XYZKX" }”
- user message : the end user question
So I get back the json , I can pass it to an API , get relevant information , such as the url where the user con navigate with a deep link … build some text accordingly
I sent to chat gpt
- user message : the reply from the API
- user message : the initial end user question
first tests looks promising ,. still I have some doubts such has
-
is it a viable apporach or an anti pattern ?
-
Currently I am working on a chatbot that is already working fine to provide domain specific information (I am adding conversation history and relevant documentation information from embeddings) . I am thinking to add this intent detector / resolver in the chain , but not sure if I should drop conversation / content embedding from the question to chatgpt if I have detected / resolved an intent in the processing chain . Any advice ?
