Dynamic configuration of NLU endpoint (Tracker Aware Interpreter)

Enhancement proposal: The interpreter invokes the NLU service to parse each message. The service endpoint URL is configurable via a configuration file and cannot be dynamically changed at runtime. The NLU endpoint receives the text of the message but doesn’t receive the representation of the conversation.

The proposal is to allow the support of the following conceptual use cases:

1. Greeting in the right language

• The user signs up to a chat app, specifies their communication preferences (eg. English or French)
• Rasa utters the message “Hello!” or “Bonjour!” depending on the communication preference

1. Language-specific NLU parsing

• The user signs up to a chat app, specifies their communication preferences (eg. English or French)
• The user starts the chat by sending a message in a language of their choice
• Rasa utters the response in the same language

1. Multi-language interaction

• The user starts the chat in French
• Responses are in French
• The user sends a message in English
• Rasa sends the response to English

1. Co-reference resolution Restaurant reservation chatbot Rasa: How many people joining you for dinner? User: Just one Rasa: Ok, what is your meal preference? User: Vegetarian Rasa: Great choice, what about the other party? User: Same as me Rasa: Ok, confirmed. Party 1: vegetarian; Party 2: vegetarian

The task here is to determine that the message “Same as me” refers to the previous message.

Proposed solution

• Enhance the interpreter to have access to the tracker.
• Enhance the NLU configuration to specify the slot name that will contain the URL of the NLU service endpoint
• Enhance parse API of the NLU service to accept tracker state as input

[alexissmirnov]

Exactly, @amn41

If the tracker is passed into the interpreter, it would open all this logic/flexibility

[amn41]

If I’ve understood correctly, @alexissmirnov - the problem is that in processor.py only the message text is passed to the interpreter class, correct? because otherwise you could create an interpreter subclass which does all of this custom logic (e.g. language detection, looking at a slot, using different NLU models per language, etc)

https://github.com/RasaHQ/rasa_core/blob/978c76a253e5d5a8e8dc61521b5589dc94f28b08/rasa_core/processor.py#L263