Hi, I haven’t really worked on that because 1. The training data will be different from standard Rasa NLU format and more similar to Rasa stories instead, 2. Since the sentence that requires context for chatbot might be quite rare, I think for now it’s better to use the end-to-end story instead.
If you still want to do this, as I said, I think you can create custom NaturalLanguageInterpreter and Interpreter class. I think you can pass the tracker here, put the metadata here, and you need to replace agent.interpreter as well.