actions.rst

desc:	Learn about about how to write your own custom actions with the open source Rasa framework to be able to interact with the external world - ranging from databases to third-party APIs.

Actions

.. edit-link::

Actions are the things your bot runs in response to user input. There are four kinds of actions in Rasa:

1. Utterance actions: start with utter_ and send a specific message to the user
2. Retrieval actions: start with respond_ and send a message selected by a retrieval model
3. Custom actions: run arbitrary code and send any number of messages (or none).
4. Default actions: e.g. action_listen, action_restart, action_default_fallback

• Utterance Actions
• Retrieval Actions
• Custom Actions
  • Custom Actions Written in Python
• Execute Actions in Other Code
• Proactively Reaching Out to the User Using Actions
• Default Actions

Utterance Actions

To define an utterance action (ActionUtterTemplate), add an utterance template to the domain file that starts with utter_:

templates:
  utter_my_message:
    - "this is what I want my action to say!"

It is conventional to start the name of an utterance action with utter_. If this prefix is missing, you can still use the template in your custom actions, but the template can not be directly predicted as its own action. See :ref:`responses` for more details.

If you use an external NLG service, you don't need to specify the templates in the domain, but you still need to add the utterance names to the actions list of the domain.

Retrieval Actions

Retrieval actions make it easier to work with a large number of similar intents like chitchat and FAQs. See :ref:`retrieval-actions` to learn moree.

Custom Actions

An action can run any code you want. Custom actions can turn on the lights, add an event to a calendar, check a user's bank balance, or anything else you can imagine.

Rasa will call an endpoint you can specify, when a custom action is predicted. This endpoint should be a webserver that reacts to this call, runs the code and optionally returns information to modify the dialogue state.

To specify, your action server use the endpoints.yml:

action_endpoint:
  url: "http://localhost:5055/webhook"

And pass it to the scripts using --endpoints endpoints.yml.

You can create an action server in node.js, .NET, java, or any other language and define your actions there - but we provide a small python SDK to make development there even easier.

Note

Rasa uses a ticket lock mechanism to ensure incoming messages from the same conversation ID do not interfere with each other and are processed in the right order. If you expect your custom action to take more than 60 seconds to run, please set the TICKET_LOCK_LIFETIME environment variable to your expected value.

Custom Actions Written in Python

For actions written in python, we have a convenient SDK which starts this action server for you.

The only thing your action server needs to install is rasa-sdk:

pip install rasa-sdk

Note

You do not need to install rasa for your action server. E.g. it is recommended to run Rasa in a docker container and create a separate container for your action server. In this separate container, you only need to install rasa-sdk.

The file that contains your custom actions should be called actions.py. Alternatively, you can use a package directory called actions or else manually specify an actions module or package with the --actions flag.

If you have rasa installed, run this command to start your action server:

rasa run actions

Otherwise, if you do not have rasa installed, run this command:

python -m rasa_sdk --actions actions

In a restaurant bot, if the user says "show me a Mexican restaurant", your bot could execute the action ActionCheckRestaurants, which might look like this:

.. testcode::

   from rasa_sdk import Action
   from rasa_sdk.events import SlotSet

   class ActionCheckRestaurants(Action):
      def name(self) -> Text:
         return "action_check_restaurants"

      def run(self,
              dispatcher: CollectingDispatcher,
              tracker: Tracker,
              domain: Dict[Text, Any]) -> List[Dict[Text, Any]]:

         cuisine = tracker.get_slot('cuisine')
         q = "select * from restaurants where cuisine='{0}' limit 1".format(cuisine)
         result = db.query(q)

         return [SlotSet("matches", result if result is not None else [])]

You should add the the action name action_check_restaurants to the actions in your domain file. The action's run method receives three arguments. You can access the values of slots and the latest message sent by the user using the tracker object, and you can send messages back to the user with the dispatcher object, by calling dispatcher.utter_message().

Details of the run() method:

.. automethod:: rasa_sdk.Action.run

There is an example of a SlotSet event :ref:`above <custom_action_example>`, and a full list of possible events in :ref:`Events <events>`.

Execute Actions in Other Code

Rasa will send an HTTP POST request to your server containing information on which action to run. Furthermore, this request will contain all information about the conversation. :ref:`action-server` shows the detailed API spec.

As a response to the action call from Rasa, you can modify the tracker, e.g. by setting slots and send responses back to the user. All of the modifications are done using events. There is a list of all possible event types in :ref:`events`.

Proactively Reaching Out to the User Using Actions

You may want to proactively reach out to the user, for example to display the output of a long running background operation or notify the user of an external event.

To do so, you can POST to this endpoint , specifying the action which should be run for a specific user in the request body. Use the output_channel query parameter to specify which output channel should be used to communicate the assistant's responses back to the user. If your message is static, you can define an utter_ action in your domain file with a corresponding template. If you need more control, add a custom action in your domain and implement the required steps in your action server. Any messages which are dispatched in the custom action will be forwarded to the specified output channel.

Proactively reaching out to the user is dependent on the abilities of a channel and hence not supported by every channel. If your channel does not support it, consider using the :ref:`callbackInput` channel to send messages to a webhook.

Note

Running an action in a conversation changes the conversation history and affects the assistant's next predictions. If you don't want this to happen, make sure that your action reverts itself by appending a ActionReverted event to the end of the conversation tracker.

Default Actions

There are eight default actions:

action_listen	Stop predicting more actions and wait for user input.
action_restart	Reset the whole conversation. Can be triggered during a conversation by entering /restart if the :ref:`mapping-policy` is included in the policy configuration.
action_session_start	Start a new conversation session. Take all set slots, mark the beginning of a new conversation session and re-apply the existing SlotSet events. This action is triggered automatically after an inactivity period defined by the session_expiration_time parameter in the domain's session config. Can be triggered manually during a conversation by entering /session_start.
action_default_fallback	Undo the last user message (as if the user did not send it and the bot did not react) and utter a message that the bot did not understand. See :ref:`fallback-actions`.
action_deactivate_form	Deactivate the active form and reset the requested slot. See also :ref:`section_unhappy`.
action_revert_fallback_events	Revert events that occurred during the TwoStageFallbackPolicy. See :ref:`fallback-actions`.
action_default_ask_affirmation	Ask the user to affirm their intent. It is suggested to overwrite this default action with a custom action to have more meaningful prompts.
action_default_ask_rephrase	Ask the user to rephrase their intent.
action_back	Undo the last user message (as if the user did not send it and the bot did not react). Can be triggered during a conversation by entering /back if the MappingPolicy is included in the policy configuration.

All the default actions can be overwritten. To do so, add the action name to the list of actions in your domain:

actions:
- action_default_ask_affirmation

Rasa will then call your action endpoint and treat it as every other custom action.