Proxifing rasa

Hello. In my application I want to use RASA not directly from UI, but from my Java microservice. User can ask questions which have rasa intents (e.g. what is the weather like today), but also some other questions which should be handled separately (not in rasa but in the java application, e.g.: What is the highest mountain in the world?). So I call HTTP API from my java application to get user intent and verify the intent and confidence. If the confidence is ok (e.g. confidence > 0.8) I do appropriate action for the intent and return answer to the user. But I have 3 problems:

  1. How should I implement the stories in such case? For example: H: what is the weather like today? B: In what city? H: In Paphos B: Sunny
  2. Do my calls to the addConversationMessage affect the response of the subsequent calls to the same endpoint. Because, if I consume intent from addConversationMessage response and do appropriate action, the next intent from addConversationMessage response shouldn’t be affected by the consumed.
  3. Is it reasonable implementation just to check the confidence of intent and execute another action not related to rasa? Or should I train the rasa model on a lot of various questions as a negative examples and assign some “qa” intent to it and check for this qa intent instead of checking the confidence?