Skip to content
Input validation, disambiguation, and rules on top of Rasa Core
Branch: master
Clone or download
Latest commit 26abc60 Feb 11, 2019
Type Name Latest commit message Commit time
Failed to load latest commit information.
.github/ISSUE_TEMPLATE port to 0.11 Sep 27, 2018
rasa_addons Merge branch 'swap_none_intents' Jan 16, 2019
tests after rule for None intent Dec 4, 2018
.gitignore Merge branch 'master' of Jul 31, 2018
CHANGELOG.rst changelog Jan 16, 2019
LICENSE init Apr 9, 2018
requirements.txt Minor version bump Jan 17, 2019

Rasa Addons

PyPI Travis

A set of power tools to 🚀🚀🚀 your productivity with Rasa

  • Input validation: if you expect Yes or No, make sure your users answer Yes or No
  • Disambiguation and fallback: automatically display dismabiguation options to users based on custom triggers
  • Intent Substitution: avoid random intents when users enter data without semantic consistency (names, brands, time,...)
  • Filter entities: define entities allowed for each intent



Rasa core < 0.11.x

pip install rasa-addons==0.4.3

Rasa core >= 0.11.x

pip install rasa-addons


You can set rules in a declarative way using a YAML file or a remote endpoint. To do that you must start Rasa Core from a python script and include the following snippet

To load rules from a YAML file:

from rasa_addons.superagent import SuperAgent
agent = SuperAgent.load(...,rules='rules.yml')

To load rules from a remote endpoint:

from rasa_addons.superagent import SuperAgent
agent = SuperAgent.load(...,rules=EndpointConfig(url="https://my.rules.endpoint/path", ...))

In the rest of this document we'll assume you are reading from a YAML file

Validate user input

from rasa_addons.superagent import SuperAgent
agent = SuperAgent.load(...,rules='rules.yml')

In rules.yml you can add input validation rules

  - after: utter_when_do_you_want_a_wake_up_call
    # !!WARNING!! If regex is set true then the validation will trigger for
    #             all actions which includes the above text. It is encouraged
    #             to set regex to false for matching the validation to a
    #             specific action.
    regex: false # optional (default: True)
      - intents:
        - cancel
      - intents:
        - skeak_to_human
      - intents:
        - enter_time
        - time
    error_template: utter_please_provide_time

The following rule will utter the error_template if the user does not reply to utter_when_do_you_want_a_wake_up_call with either /cancel OR /speak_to_human OR /enter_time{"time":"..."} Rules are enforced at the tracker level, so there is no need to retrain when changing them.

Disambiguate user input and fallback

Disambiguation policy

Help your users when your NLU struggles to identify the right intent. Instead of just going with the highest scoring intent or just going with a fallback you can ask the user to confirm the question or to pick from a list of likely intents.


One way to disambiguate is to provide the user with buttons, each button corresponding to one intent. In the example below, the disambiguation is triggered when the score of the highest scoring intent is below twice the score of the second highest scoring intent.

The bot will utter:

  1. An intro message (if the optional field intro_template is present)
  2. A text with buttons (or quick replies) message where:
  • the text is the template defined as text_template,
  • the button titles will be the concatenation of "utter_disamb" and the intent name. For example, utter_disamb_greet."
  • the buttons payloads will be the corresponding intents (e.g. /greet). Entities found in parse_data are passed on.
  1. A fallback button to go along with disambiguation buttons (if the optional field fallback_button is present)

It's also possible to exclude certain intents from being displayed as a disambiguation option by using optional exclude list field. In the example below, all intents that match regex chitchat\..* and basics\..*, as well as intent cancel will not be displayed as an option. The next highest scoring intents will be displayed in place of excluded ones.

  trigger: $0 < 2 * $1
  type: suggest
  max_suggestions: 2
  slot_name: parse_data # optional slot name to store the parse data originating a disambiguation
    intro_template: utter_disamb_intro # optional: will not be rendered if not set
    text_template: utter_disamb_text
    button_title_template_prefix: utter_disamb
      title: utter_fallback_yes
      payload: /fallback
      - chitchat\..*
      - basics\..*
      - cancel


  • trigger: $0 corresponds to parse_data['intent_ranking'][0]["confidence"]. You can set any rule based on intent ranking. Intent scores are checked against the trigger before any intent is excluded with exclude.
  • slot_name: you need to set the slot in the Core domain to get it from the tracker. E.g. tracker.get_slot(slot_name)You may want to make the bot go straight to suggesting fallback (e.g when the top intent ranking is low).

The bot will utter:

  1. An intro message utter_fallback_intro
  2. Optional buttons (if buttons list with at least one item - a pair of title and payload - is defined).


Another way to disambiguate is to rephrase. When triggered, the bot asks "Did you mean [something related to the intent]"? followed by two buttons (titles in yes_template and no_template). no_payload is the payload to trigger when the user clicks the no button.

  trigger: $0 < 2 * $1
  type: rephrase
    rephrase_template: utter_rephrase
    yes_template: utter_yes
    no_template: utter_no
    no_payload: /fallback
      - chitchat\..*
      - basics\..*
      - cancel

Fallback policy

In the example below, fallback is triggered when the top scoring intent's confidence is below 0.5.

  trigger: $0 < 0.5
  slot_name: parse_data # optional slot name to store the parse data originating a disambiguation
    text: utter_fallback_intro
      - title: utter_fallback_yes
        payload: /fallback
      - title: utter_fallback_no
        payload: /restart

There is no limit on the number of buttons you can define for fallback. If no buttons are defined, this policy will simply make the bot utter some default message (e.g utter_fallback_intro) when the top intent confidence is lower than the trigger.

Using both disambiguation and fallback policies

It's easy to combine both disambiguation and fallback policies. It can be done by filling in policy definitions from two previous examples as follows:

      (...disambiguation policy definition...)

      (...fallback policy definition...)

In cases when intent confidence scores in parsed data are such that would cause both policies to trigger, only fallback policy is trigerred. In other words, fallback policy has precedence over disambiguation policy.

Substitute intents

Some intents are hard to catch. For example when the user is asked to fill arbitrary data such as a date or a proper noun. The following rule swaps any intent caught after utter_when_do_you_want_a_wake_up_call with enter_data unless...

  - after: utter_when_do_you_want_a_wake_up_call
    intent: enter_data
    unless: frustration|cancel|speak_to_human

Filter entities

Sometimes Rasa NLU CRF extractor will return unexpected entities and those can perturbate your Rasa Core dialogue model because it has never seen this particular combination of intent and entity.

This helper lets you define precisely the entities allowed for every intent in a yaml file. Entities not in the list for a given intent will be cleared. It will only remove entities for intents specifically listed in this section:

  book: # intent
    - origin # entity
    - destination
    - color
    - product

Then load your agent

agent = SuperAgent.load(POLICY_PATH,
                        create_dispatcher=lambda sender_id, output_channel, domain: MyDispatcher(sender_id, output_channel, domain))

Bonus - Create a FAQ bot with only ONE action and ONE story

You create an intent substitution rule like this:


  - intent: (faq.*)
    with: faq
      - name: intent
        value: '{intent}'

This rule will match all intents starting with faq (e.g.: faq.how_do_i_create_a_faq) This will change the dialog act to {intent: "faq", entities: [{intent: "faq.how_do_i_create_a_faq"}]}

In Core, add this story:

## FAQ
* faq{"intent":"original_intent"}
  - action_faq

And this action

class ActionFAQ(Action):

    def name(self):
        return "action_faq"

    def run(self, dispatcher, tracker, domain):
         # get the original intent from tracker.latest_message and retrieve the correct answer

The benefit of this approach is you have only ONE story for all your questions, so if your Q&A are stored externally you don't have to retrain your bot when adding/changing questions. Since you have only one story for potentially 100's of questions, this means you can better handle side questions in more complex dialogs.

Run automated tests (experimental)

You can write test cases as you would write stories, except you should only have utter_... actions.

## chitchat.greet
* chitchat.greet
  - utter_reply_to_greet

## chitchat.how_are_you
* chitchat.how_are_you
  - utter_reply_to_how_are_you

## chitchat.are_you_a_robot
* chitchat.are_you_a_robot
  - utter_reply_to_are_you_a_robot

The test session sends the user utterances via http POST requests to the rasa_core server endpoint specified in the host parameter. For this to work, make sure that your rasa core instance is running and has rest: written in the config/credentials.yml file (this tells rasa_core to accept REST api requests).

You can run the tests with the command

python -m rasa_addons.tests --host localhost:5005 -t test_cases/ # -s(--shuffle) -u(--distinct) -v(--verbose)

You can put your test cases in different files starting with test (e.g. a directory.
At this time, it only runs the test and outputs dialogues in the console (errors in red). There is no report (Help wanted). You can also use --distinct to change the sender_id at every test case and --shuffle to shuffle test cases before running the tests.

You can’t perform that action at this time.