Thanks for your answer! My problem is that the message would be evaluated as an intend. And as the input can be anything it could trigger any intend or utter_default. How am I supposed to train my NLU model that the * inform intend can be everything?
You can check how Rasa does it with a similar intent (enter_data) rasa-demo/nlu.md at master · RasaHQ/rasa-demo · GitHub
Alternatively, if you want it to be a bit more reliable you could create a simple custom policy with something like this in the predict_action_probability method (pseudo code)
def predict_action_probabilities(...params): # You can look the parameters up
result = [0.0] * domain.num_actions # domain is one of the params
slot_requested = tracker.get_slot(SLOT_NAME)
if slot_requested:
# Last action was a message by the user so we want to save it
if tracker.latest_action_name == ACTION_LISTEN_NAME:
idx = domain.index_for_action(SAVE_SLOT_ACTION)
result[idx] = 1.0
else:
# Wait for the user to send us something
idx = domain.index_for_action(ACTION_LISTEN_NAME)
result[idx] = 1.0
return result