I made a little sample for you. It starts out by using the Raw input to just answer the users message normally, this can be deleted but then you don't have a message to the user like "Hi there. What can I help you with today?". Then we wait for the users input to proceed. We then have an expression card that checks if the user wants to end the conversation, you can make it more resilient by adding more word to it. You can also use the intent card which the 2nd sample I sent uses. We then have a Query Knowledge Bases card to query the KB. We then have an expression card to check if it found an answer in the KB. If it did it will just continue down and print out the answer with the text card. If it wasn't able to find an answer to the users question it will continue over to the GPT_Fallback and use GPT-3.5 to try and answer the users question. The GPT Fallback can be replaced by a simple message saying something like I'm sorry, I wasn't able to answer that and then it could just go back to the handleUserInput node.