Rasa is the leading open-source conversational AI platform that enables both individual developers and large enterprises to build superior AI assistants and chatbots. Rasa provides the infrastructure and tools needed to build the outstanding tools and transform the way customers communicate with businesses. Rasa can be deeply customized down to levels not possible with other platforms due to the open sourced architecture and machine learning.

Rasa is used by millions of developers and small teams to program enterprise conversational AI applications.

Rasa is available in two editions: Rasa Open Source (free) and Rasa Enterprise (commercial). Both editions can be used to build voicebots with CVG.

VIER Cognitive Voice Gateway (CVG) enables access to telephony, speech-to-text (STT), text-to-speech (TTS) and contact center integration for chatbots built with Rasa. I.e. CVG makes your chatbot to a voicebot handling phone calls.

To build voice bots using Rasa and CVG, you need an account in CVG and a Rasa installation.

Rasa can be hosted anywhere: in the cloud, On-Prem or in any data center. Migrations between hosting solutions can be performed at any stage.

Many organizations developing chatbots and voicebots with Rasa start with Rasa Open Source On-Prem. An installation guide is provided by Rasa. Rasa also provides a Playground that can be used to develop bots without requiring an On-Prem installation.

To install Rasa Enterprise, use the installation guide provided by Rasa.

The Rasa integration with CVG is done with a new channel for Rasa provided in this repo. It implements all the CVG APIs relevant for bots to provide CVGs full power to you as a Rasa developer.

The easiest way to install this package is through PyPI.

pip install rasa-vier-cvg

If you are using Rasa on Docker and you don't want to build a derived image, you can also download the channel source and bind-mount the package into a rasa/rasa-base container with a volume definition like this:

./rasa_vier_cvg:/opt/venv/lib/python3.10/site-packages/rasa_vier_cvg

Add the following content to credentials.yml:

rasa_vier_cvg.CVGInput:
  token: "CHOOSE_YOUR_TOKEN"
  blocking_endpoints: false

You can generate the token yourself. For example with any password generator.

This channel will be used for communication with CVG. The Bot token is required so that Rasa can verify that CVG is communicating with your Rasa Bot.

The optional blocking_endpoints option allows to disable blocking CVG's request while processing the user message. For compatibility reasons this option defaults to true, but we recommend setting it to false.

If you do not yet have an account for CVG please contact us at info@vier.ai.

To configure the connection between your Rasa bot and CVG just select Rasa as the bot template, enter your Rasa URL (e.g. https://rasa.example.org/webhooks/vier-cvg) and your token, as set in credentials.yaml, in the CVG project settings.

Every message and intent sent by CVG will have a metadata-field called cvg_body. This field will always contain the JSON sent by CVG to the Rasa channel. In the following sections, the term "metadata" will refer to this cvg_body field.

Normal spoken inputs from the user as well as DTMF inputs will be transmitted as text inputs to Rasa. All other CVG events will trigger specific intents as described below. All messages and intents will have CVG's dialog ID as the sender_id field.

Text inputs follow this specification. An example for the text input metadata would be:

{
  "dialogId": "09e59647-5c77-4c02-a1c5-7fb2b47060f1",
  "projectContext": {
    "projectToken": "d30b1c38-b2fd-39c8-bec2-b268871338b0",
    "resellerToken": "ed4aff6d-c6f8-4ac9-ab67-d072ef45d9a0"
  },
  "timestamp": 1535546718115,
  "type": "SPEECH",
  "text": "Hello!",
  "confidence": 100,
  "vendor": "GOOGLE",
  "language": "en-US",
  "callback": "https://cognitivevoice.io/v1"
}

Voice and DTMF inputs can be differentiated using the type field, which would be SPEECH for voice and DTMF for DTMF tones.

Here is a list of the intents triggered by CVG for certain events:

• cvg_session: This intent is triggered once (after a new call has been established) before anything else to allow the bot to respond e.g. with a greeting. Metadata is defined by this specification
• cvg_terminated: This intent is triggered once the conversation has been terminated by the user. Metadata is specified here.
• cvg_inactivity: This intent is triggered once the inactivity timeout has been triggered due to a lack of user input. Metadata is specified here.
• cvg_recording: This intent is triggered once the recording status changes. Metadata is specified here.
• cvg_answer_number, cvg_answer_multiplechoice and cvg_answer_timeout: These intents are triggered once a prompt (see next section) of type Number or MultipleChoice complete are timeout. Metadata is specified here.
• cvg_outbound_success: The success result of forward or bridge (see next section). It signals that the outgoing call has been successfully established. Metadata is specified by the response objects of the matching operations from the Call API.
• cvg_outbound_failure: The failure result of forward or bridge (see next section). It signals that the outgoing call could not be established and provides some details as to why. Metadata is specified by the response objects of the matching operations from the Call API. Depending on the exact reason (check out the OutboundCallFailure model in the API specification for all possible reasons) there might not be a ringStartTimestamp and the ringTime could be zero.

The output channel for CVG supports text_messages and custom_json.

Text messages will be translated into Say-commands.

Every other command supported by the channel must be triggered by using custom JSON. The key for the custom JSON messages is an encoding of CVG's API endpoints and follows this schema:

cvg_<path with underscores instead of slashes>

So for example in order to use the /call/play endpoint you would use cvg_call_play as the key, for /call/transcription/switch it would be cvg_call_transcription_switch and so on.

The JSON values will be used as-is as the request-body for the API call, so refer to the API documentations, most commonly the Call API for specifics. The only exception to this is, that the dialog ID (sender_id) which is automatically injected into the payloads as necessary.

Currently all operation documented in the Call API as well as dialog_delete and dialog_data are implemented.

In case you want to call an API endpoint which is a bit more complex like /call/forward or something that is currently not implemented in this channel, you can use simply make the request manually using python.

After setting up your Rasa Installation and configuring the CVG Project, let's create a simple Rasa Bot together.
Create a new folder and generate the default bot:

rasa init

The bot is ready to be tested. Make sure you expose it in a way CVG can reach it, and configure the CVG channel.

You can start the Rasa bot using rasa run. Make sure, you run rasa train after modifying the bot.

Please paste the following intents into your domain.yml. See below, on how the intent section should look like. They are explained above, but don't worry about that yet.

intents:
  - greet
  - goodbye
  - affirm
  - deny
  - mood_great
  - mood_unhappy
  - bot_challenge

  - cvg_outbound_success
  - cvg_outbound_failure
  - cvg_session
  - cvg_answer_multiplechoice # you can remove cvg_answer_*, if you don't use the /call/prompt feature.
  - cvg_answer_number
  - cvg_answer_timeout
  - cvg_message
  - cvg_inactivity
  - cvg_terminated
  - cvg_recording

To end the call / hang up after it said "Bye", you can modify the utter_goodbye message in the domain.yml like this:

  utter_goodbye:
  - text: "Bye"
    custom:
      cvg_call_drop:

To forward the caller to an agent, you can modify utter_iamabot like this:

  utter_iamabot:
  - text: "I am a bot, powered by Rasa. But I will gladly forward you to a human."
    custom:
      cvg_call_forward:
        destinationNumber: "+4969907362380"

Please reference the Rasa documentation, on how to create and call a custom action. This example will use the default Rasa action server, which you can start with rasa run actions

class ActionPrintCvgBody(Action):

    def name(self) -> Text:
        return "action_print_cvg_body" # The action name which you can use in your domain.yml

    def run(self, dispatcher: CollectingDispatcher, tracker: Tracker, domain: Dict[Text, Any]) -> List[Dict[Text, Any]]:
        try:
          cvg_body = None
          for e in tracker.events[::-1]: # The loop will find the last message from the user
            if e["event"] == "user":
              cvg_body = e["metadata"]["cvg_body"]
              break
          print("Found cvg_body: ", cvg_body) # After we found the last message from the user and stored the CVG response body in cvg_body, we can print it
        except KeyError as e:
          print("Failed to read cvg_body: ", e) # The last user message did not contain the cvg_body. 
          # Note: The cvg_body is added by the CVG channel and won't be available if you use a different channel
        finally:
          return []

You may already notice that the bot immediately says something after calling. That is because we haven't told Rasa yet how to handle the cvg_session intent.
That intent is triggered when /session in the Bot API is called.

In your stories.yml replace the intent greet with cvg_session:

  steps:
-   - intent: greet
+   - or:
+    - intent: greet
+    - intent: cvg_session

Make sure to do that with all 3 stories and run rasa train before starting the Rasa bot.

To extract more information from the message inside an action, please read about Events above.

The intents cvg_outbound_success and cvg_outbound_failure are relevant if you want to forward or bridge a call.
You could do something like this in your domain.yml:

  utter_outbound_failure:
  - text: "Unfortunatly, the outbound call failed."

and in your rules.yml:

- rule: Handle outbound call failure
  steps:
  - intent: cvg_outbound_failure
  - action: utter_outbound_failure

This will inform the user about outbound call failures. To handle the cvg_outbound_success intent, you can create an action, but we cannot say something to a call that has already been forwarded.

If you want to use the /call/prompt feature to prompt for a number, you can create the prompt and responses in your domain.yml:

  utter_prompt:
    - custom:
      cvg_call_prompt:
        text: Please provide 3 Numbers
        timeout: 10000
        type: 
          name: Number
          maxDigits: 3
          submitInputs:
            - DTMF_#

  utter_prompt_answer_number:
  - text: "You can access the result of the prompt inside a custom action."

  utter_prompt_timeout:
  - text: "You did not provide an answer, the prompt timed out"

For how the write such an action, see below.

And add the following rules inside your data/rules.yml:

- rule: Handle prompt timeout
  steps:
  - intent: cvg_answer_timeout
  - action: utter_prompt_timeout

- rule: Handle prompt answer
  steps:
  - intent: cvg_answer_number
  - action: utter_prompt_answer_number

We provide a demo voicebot built with Rasa and CVG on GitHub. We also run this voicebot, so you can simply get a first impression. For more information, visit our GitHub project.

• When CVG sends an event to Rasa, this channel will generate the intent (as specified above)
  • The intent's metadata will contain the body sent by CVG as specified in the Bot API
• When you utter a text message, or a Custom Response, we pass the content of the payload to CVG after adding the dialog_id