Warning: This document is for an old version of Rasa. The latest version is 1.3.9.


If you want your assistant to respond to user messages, you need to manage these responses. In the training data for your bot, your stories, you specify the actions your bot should execute. These actions can use utterances to send messages back to the user.

There are two ways to manage these utterances:

  1. Include your assistant utterances in your domain file, or
  2. Use an external service to generate the responses.

Including the utterances in the domain

The default format is to include the utterances in your domain file. This file then contains references to all your custom actions, available entities, slots and intents.

# all hashtags are comments :)
 - greet
 - default
 - goodbye
 - affirm
 - thank_you
 - change_bank_details
 - simple
 - hello
 - why
 - next_intent

 - name

    type: text

    - text: "hey there {name}!"  # {name} will be filled by slot (same name) or by custom action
    - text: "this is a default channel"
    - text: "you're talking to me on slack!"  # if you define channel-specific utterances, the bot will pick
      channel: "slack"                        # from those when talking on that specific channel
    - text: "goodbye 😢"   # multiple templates - bot will randomly pick one of them
    - text: "bye bye 😢"
  utter_default:   # utterance sent by action_default_fallback
    - text: "sorry, I didn't get that, can you rephrase it?"

  - utter_default
  - utter_greet
  - utter_goodbye

In this example domain file, the section templates contains the template the assistant uses to send messages to the user.

If you want to change the text, or any other part of the bots response, you need to retrain the assistant before these changes will be picked up.

More details about the format of these responses can be found in the documentation about the domain file format: Utterance templates.

Managing assistant utterances using an external CMS

Retraining the bot just to change the text copy can be suboptimal for some workflows. That’s why Core also allows you to outsource the response generation and separate it from the dialogue learning.

The assistant will still learn to predict actions and to react to user input based on past dialogues, but the responses it sends back to the user are generated outside of Rasa Core.

If the assistant wants to send a message to the user, it will call an external HTTP server with a POST request. To configure this endpoint, you need to create an endpoints.yml and pass it either to the run or server script. The content of the endpoints.yml should be

  url: http://localhost:5055/nlg    # url of the nlg endpoint
  # you can also specify additional parameters, if you need them:
  # headers:
  #   my-custom-header: value
  # token: "my_authentication_token"    # will be passed as a get parameter
  # basic_auth:
  #   username: user
  #   password: pass
# example of redis external tracker store config
  type: redis
  url: localhost
  port: 6379
  db: 0
  password: password
  record_exp: 30000
# example of mongoDB external tracker store config
  #type: mongod
  #url: mongodb://localhost:27017
  #db: rasa
  #user: username
  #password: password

Then pass the enable-api flag to the rasa run command when starting the server:

$ rasa run \
   --enable-api \
   -m examples/babi/models \
   --log-file out.log \
   --endpoints endpoints.yml

The body of the POST request sent to the endpoint will look like this:

  "tracker": {
    "latest_message": {
      "text": "/greet",
      "intent_ranking": [
          "confidence": 1.0,
          "name": "greet"
      "intent": {
        "confidence": 1.0,
        "name": "greet"
      "entities": []
    "sender_id": "22ae96a6-85cd-11e8-b1c3-f40f241f6547",
    "paused": false,
    "latest_event_time": 1531397673.293572,
    "slots": {
      "name": null
    "events": [
        "timestamp": 1531397673.291998,
        "event": "action",
        "name": "action_listen"
        "timestamp": 1531397673.293572,
        "parse_data": {
          "text": "/greet",
          "intent_ranking": [
              "confidence": 1.0,
              "name": "greet"
          "intent": {
            "confidence": 1.0,
            "name": "greet"
          "entities": []
        "event": "user",
        "text": "/greet"
  "arguments": {},
  "template": "utter_greet",
  "channel": {
    "name": "collector"

The endpoint then needs to respond with the generated response:

    "text": "hey there",
    "buttons": [],
    "image": null,
    "elements": [],
    "attachments": []

Rasa will then use this response and sent it back to the user.