Port existing custom chatbot as Google Assistant action

Question

We have a framework that implements chatbot / voice assistant logic for handling complex conversations in the health domain. Everything is implemented on our server side. This gives us full control of how responses are generated.

The channel (such as Alexa or Facebook Messenger cloud) calls our webhook:

• When user messages, the platform sends these to our webhook: hashed user id, message text (chat message or transcribed voice)
• Our webhook responds with the appropriately structured response, which includes text to be displayed, spoken, possibly choice buttons and some images etc. It also includes a flag whether the current session has finished or user input is expected.

Integrating a new channel involves conversion of the response returned into the form expected by a channel and setting some flags (has voice, has display etc.).

This simple framework has worked so far for Facebook Messenger, Cortana, Alexa (a little bit of hacking was needed to abandon it's intent and slot recognition), our web chatbot .

We wanted to write a thin layer of support for Google Assistant action.

Is there any way of passing all the input from Assistant user intact into a webhook such as the one described above and taking full control of the way responses are generated and the end of conversation is determined?

I'd rather not delve into those cumbersome ways of API.AI of structuring a conversation which seems good for a trivial scenarios such as ordering an Uber but seems very bad for longer conversation.

Answer 1

Since you already have a Natural Language Understanding layer for your system, you don't need API.AI/Dialogflow, and you can skip this layer completely. (The NLU is useful, even for large and extensive conversations, but doesn't make sense in your case where you've already defined the conversation through other means.)

You'll need to use the Actions SDK (sometimes known as actions.json after the configuration file it uses) to define triggering phrases, but after that you'll get all the text that the user says as part of your conversation through a webhook that delivers JSON to you. You'll reply with JSON that contains the text/audio response, images on cards, possibly suggestion chips, etc.