Intent classification and entity recognition


I’d like to know if Rasa classify the intent and extract entity at the same time, meaning, the output include both the intent classification and potential entities, or it does it separately, meaning, the model first classify the intent, then if the predicted intent include entities within the training data, then another model (SpaCy or Duckling) gets called and then extract the entities (or maybe it does it anyway, whether there is entities on it or not).

If it does it separately, does this mean only the intent classification is retrained when needed, while the entity extraction model is set once and then re-used when needed (doesn’t necessarily need re-train if it does the work).