Tutorial: Building Assistants

After following the basics of setting up an assistant in the Rasa Tutorial, we’ll now walk through building a basic FAQ chatbot and then build a bot that can handle contextual conversations.

Building a simple FAQ assistant

FAQ assistants are the simplest assistants to build and a good place to get started. These assistants allow the user to ask a simple question and get a response. We’re going to build a basic FAQ assistant using features of Rasa designed specifically for this type of assistant.

In this section we’re going to cover the following topics:

• Responding to simple intents with the MemoizationPolicy
• Handling FAQs using the ResponseSelector

You should first install Rasa using the Step-by-step Installation Guide and then follow the Rasa Tutorial to make sure you know the basics.

To prepare for this tutorial, we’re going to create a new directory and start a new Rasa project.

mkdir rasa-assistant
rasa init

Let’s remove the default content from this bot, so that the nlu.md, stories.md and domain.yml files are empty.

Memoization Policy

The MemoizationPolicy remembers examples from training stories for up to a max_history of turns. The number of “turns” includes messages the user sent, and actions the assistant performed. For the purpose of a simple, context-less FAQ bot, we only need to pay attention to the last message the user sent, and therefore we’ll set that to 1.

You can do this by editing your config.yml file as follows:

policies:
- name: MemoizationPolicy
  max_history: 1
- name: MappingPolicy

Now that we’ve defined our policies, we can add some stories for the goodbye, thank and greet intents to the stories.md file:

## greet
* greet
  - utter_greet

## thank
* thank
  - utter_noworries

## goodbye
* bye
  - utter_bye

We’ll also need to add the intents, actions and responses to our domain.yml file in the following sections:

intents:
  - greet
  - bye
  - thank

responses:
  utter_noworries:
    - text: No worries!
  utter_greet:
    - text: Hi
  utter_bye:
    - text: Bye!

Finally, we’ll copy over some NLU data from Sara into our nlu.md file:

## intent:greet
- Hi
- Hey
- Hi bot
- Hey bot
- Hello
- Good morning
- hi again
- hi folks

## intent:bye
- goodbye
- goodnight
- good bye
- good night
- see ya
- toodle-oo
- bye bye
- gotta go
- farewell

## intent:thank
- Thanks
- Thank you
- Thank you so much
- Thanks bot
- Thanks for that
- cheers

You can now train a first model and test the bot, by running the following commands:

rasa train
rasa shell

This bot should now be able to reply to the intents we defined consistently, and in any order.

For example:

While it’s good to test the bot interactively, we should also add end-to-end test cases that can later be included as part of a CI/CD system. End-to-end test conversations include NLU data, so that both components of Rasa can be tested. The file tests/conversation_tests.md contains example test conversations. Delete all the test conversations and replace them with some test conversations for your assistant so far:

## greet + goodbye
* greet: Hi!
  - utter_greet
* bye: Bye
  - utter_bye

## greet + thanks
* greet: Hello there
  - utter_greet
* thank: thanks a bunch
  - utter_noworries

## greet + thanks + goodbye
* greet: Hey
  - utter_greet
* thank: thank you
  - utter_noworries
* bye: bye bye
  - utter_bye

To test our model against the test file, run the command:

rasa test --stories tests/conversation_tests.md

The test command will produce a directory named results. It should contain a file called failed_stories.md, where any test cases that failed will be printed. It will also specify whether it was an NLU or Core prediction that went wrong. As part of a CI/CD pipeline, the test option --fail-on-prediction-errors can be used to throw an exception that stops the pipeline.

Response Selectors

The ResponseSelector NLU component is designed to make it easier to handle dialogue elements like Small Talk and FAQ messages in a simple manner. By using the ResponseSelector, you only need one story to handle all FAQs, instead of adding new stories every time you want to increase your bot’s scope.

People often ask Sara different questions surrounding the Rasa products, so let’s start with three intents: ask_channels, ask_languages, and ask_rasax. We’re going to copy over some NLU data from the Sara training data into our nlu.md. It’s important that these intents have an faq/ prefix, so they’re recognised as the faq intent by the ResponseSelector:

## intent: faq/ask_channels
- What channels of communication does rasa support?
- what channels do you support?
- what chat channels does rasa uses
- channels supported by Rasa
- which messaging channels does rasa support?

## intent: faq/ask_languages
- what language does rasa support?
- which language do you support?
- which languages supports rasa
- can I use rasa also for another laguage?
- languages supported

## intent: faq/ask_rasax
- I want information about rasa x
- i want to learn more about Rasa X
- what is rasa x?
- Can you tell me about rasa x?
- Tell me about rasa x
- tell me what is rasa x

Next, we’ll need to define the responses associated with these FAQs in a new file called responses.md in the data/ directory:

## ask channels
* faq/ask_channels
  - We have a comprehensive list of [supported connectors](/content/docs/core/connectors/index.html), but if you don't see the one you're looking for, you can always create a custom connector by following [this guide](/content/docs/rasa/user-guide/connectors/custom-connectors/index.html).

## ask languages
* faq/ask_languages
  - You can use Rasa to build assistants in any language you want!

## ask rasa x
* faq/ask_rasax
 - Rasa X is a tool to learn from real conversations and improve your assistant. Read more [here](/content/docs/rasa-x/index.html)

The ResponseSelector should already be at the end of the NLU pipeline in our config.yml:

language: en
pipeline:
  - name: WhitespaceTokenizer
  - name: RegexFeaturizer
  - name: LexicalSyntacticFeaturizer
  - name: CountVectorsFeaturizer
  - name: CountVectorsFeaturizer
    analyzer: "char_wb"
    min_ngram: 1
    max_ngram: 4
  - name: DIETClassifier
    epochs: 100
  - name: EntitySynonymMapper
  - name: ResponseSelector
    epochs: 100

Now that we’ve defined the NLU side, we need to make Core aware of these changes. Open your domain.yml file and add the faq intent:

intents:
  - greet
  - bye
  - thank
  - faq

We’ll also need to add a retrieval action, which takes care of sending the response predicted from the ResponseSelector back to the user, to the list of actions. These actions always have to start with the respond_ prefix:

actions:
  - respond_faq

Next we’ll write a story so that Core knows which action to predict:

## Some question from FAQ
* faq
    - respond_faq

This prediction is handled by the MemoizationPolicy, as we described earlier.

After all of the changes are done, train a new model and test the modified FAQs:

rasa train
rasa shell

At this stage it makes sense to add a few test cases to your test_stories.md file again:

## ask channels
* faq: What messaging channels does Rasa support?
  - respond_faq

## ask languages
* faq: Which languages can I build assistants in?
  - respond_faq

## ask rasa x
* faq: What’s Rasa X?
  - respond_faq

Using the features we described in this tutorial, you can easily build a context-less assistant. When you’re ready to enhance your assistant with context, check out Building a contextual assistant.

Handling business logic

A lot of conversational assistants have user goals that involve collecting a bunch of information from the user before being able to do something for them. This is called slot filling. For example, in the banking industry you may have a user goal of transferring money, where you need to collect information about which account to transfer from, whom to transfer to and the amount to transfer. This type of behavior can and should be handled in a rule-based way, as it is clear how this information should be collected.

For this type of use case, we can use Forms and our FormPolicy. The FormPolicy works by predicting the form as the next action until all information is gathered from the user.

As an example, we will build out the SalesForm from Sara. The user wants to contact our sales team, and for this we need to gather the following pieces of information:

• Their job
• Their bot use case
• Their name
• Their email
• Their budget
• Their company

We will start by defining the SalesForm as a new class in the file called actions.py. The first method we need to define is the name, which like in a regular Action returns the name that will be used in our stories:

from rasa_sdk.forms import FormAction

class SalesForm(FormAction):
    """Collects sales information and adds it to the spreadsheet"""

def name(self):
        return "sales_form"

Next we have to define the required_slots method which specifies which pieces of information to ask for, i.e. which slots to fill.

@staticmethod
def required_slots(tracker):
    return [
        "job_function",
        "use_case",
        "budget",
        "person_name",
        "company",
        "business_email",
        ]

Once you’ve done that, you’ll need to specify how the bot should ask for this information. This is done by specifying utter_ask_{slotname} responses in your domain.yml file. For the above we’ll need to specify the following:

utter_ask_business_email:
  - text: What's your business email?
utter_ask_company:
  - text: What company do you work for?
utter_ask_budget:
  - text: "What's your annual budget for conversational AI? 💸"
utter_ask_job_function:
  - text: "What's your job? 🕴"
utter_ask_person_name:
  - text: What's your name?
utter_ask_use_case:
  - text: What's your use case?

We’ll also need to define all these slots in our domain.yml file:

slots:
  company:
    type: unfeaturized
  job_function:
    type: unfeaturized
  person_name:
    type: unfeaturized
  budget:
    type: unfeaturized
  business_email:
    type: unfeaturized
  use_case:
    type: unfeaturized

Going back to our Form definition, we need to define the submit method as well, which will do something with the information the user has provided once the form is complete:

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

dispatcher.utter_message("Thanks for getting in touch, we’ll contact you soon")
    return []

In this case, we only tell the user that we’ll be in touch with them; however, usually you would send this information to an API or a database. See the rasa-demo for an example of how to store this information in a spreadsheet.

We’ll need to add the form we just created to a new section in our domain.yml file:

forms:
  - sales_form

We also need to create an intent to activate the form, as well as an intent for providing all the information the form asks the user for. For the form activation intent, we can create an intent called contact_sales. Add the following training data to your nlu file:

## intent:contact_sales
- I wanna talk to your sales people.
- I want to talk to your sales people
- I want to speak with sales
- Sales
- Please schedule a sales call
- Please connect me to someone from sales
- I want to get in touch with your sales guys
- I would like to talk to someone from your sales team
- sales please

We will also create an intent called inform which covers any sort of information the user provides to the bot. The reason we put all this under one intent is because there is no real intent behind providing information, only the entity is important. Add the following data to your NLU file:

## intent:inform
- [100k](budget)
- [100k](budget)
- [240k/year](budget)
- [150,000 USD](budget)
- I work for [Rasa](company)
- The name of the company is [ACME](company)
- company: [Rasa Technologies](company)
- it's a small company from the US, the name is [Hooli](company)
- it's a tech company, [Rasa](company)
- [ACME](company)
- [Rasa Technologies](company)
- [maxmeier@firma.de](business_email)
- [bot-fan@bots.com](business_email)
- [maxmeier@firma.de](business_email)
- [bot-fan@bots.com](business_email)
- [my email is email@rasa.com](business_email)
- [engineer](job_function)
- [brand manager](job_function)
- [marketing](job_function)
- [sales manager](job_function)
- [growth manager](job_function)
- [CTO](job_function)
- [CEO](job_function)
- [COO](job_function)
- [John Doe](person_name)
- [Jane Doe](person_name)
- [Max Mustermann](person_name)
- [Max Meier](person_name)
- We plan to build a [sales bot](use_case) to increase our sales by 500%.
- we plan to build a [sales bot](use_case) to increase our revenue by 100%.
- a [insurance tool](use_case) that consults potential customers on the best life insurance to choose.
- we're building a [conversational assistant](use_case) for our employees to book meeting rooms.

The intents and entities will need to be added to your domain.yml file as well:

intents:
  - greet
  - bye
  - thank
  - faq
  - contact_sales
  - inform

entities:
  - company
  - job_function
  - person_name
  - budget
  - business_email
  - use_case

A story for a form is very simple, as all the slot collection happens inside the form. Therefore it doesn’t need to be covered in your stories. You just need to write a single story showing when the form should be activated. For the sales form, add this story to your stories.md file:

## sales form
* contact_sales
    - sales_form                   <!--Run the sales_form action-->
    - form{"name": "sales_form"}   <!--Activate the form-->
    - form{"name": null}           <!--Deactivate the form-->

As a final step, let’s add the FormPolicy to our config file:

policies:
  - name: MemoizationPolicy
  - name: KerasPolicy
  - name: MappingPolicy
  - name: FormPolicy

At this point, you already have a working form, so let’s try it out. Make sure to uncomment the action_endpoint in your endpoints.yml to make Rasa aware of the action server that will run our form:

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

Then start the action server in a new terminal window:

rasa run actions

Then you can retrain and talk to your bot:

rasa train
rasa shell

This simple form will work out of the box, however you will likely want to add a bit more capability to handle different situations. One example of this is validating slots, to make sure the user provided information correctly (read more about it here).

Handling unexpected user input

All expected user inputs should be handled by the form we defined above. In real situations, the user will often behave differently. In this section we’ll go through various forms of “interjections” and how to handle them within Rasa.

The decision to handle these types of user input should always come from reviewing real conversations. You should first build part of your assistant, test it with real users, and then add what’s missing. You shouldn’t try to implement every possible edge case that you think might happen, because in the end your users may never actually behave in that way.

If you have generic interjections that should always have the same single response no matter the context, you can use the Mapping Policy to handle these. It will always predict the same action for an intent, and when combined with a forgetting mechanism, you don’t need to write any stories either.

For example, let’s say you see users having conversations where they write a greeting in the middle of a conversation, maybe because they were gone for a few minutes:

To handle these, we can add a new action for greetings in our actions.py file:

from rasa_sdk import Action
from rasa_sdk.events import UserUtteranceReverted

class ActionGreetUser(Action):
    """Revertible mapped action for utter_greet"""

def name(self):
        return "action_greet"

def run(self, dispatcher, tracker, domain):
        dispatcher.utter_template("utter_greet", tracker)
        return [UserUtteranceReverted()]

To test the modified intents, we need to re-start our action server. Then we can retrain the model, and try out our additions:

rasa train
rasa shell

FAQs are another kind of generic interjections that should always get the same response. For example, a user might ask a related FAQ in the middle of filling a form. To handle this, we can add a story for FAQ questions that could interrupt the form:

## just sales, continue
* contact_sales
    - sales_form
    - form{"name": "sales_form"}
* faq
    - respond_faq
    - sales_form
    - form{"name": null}

This will break out of the form and deal with the user's FAQ question, and then return back to the original task.

More complex contextual conversations

Not every user goal you define will fall under the category of business logic. For the other cases, you will need to use stories and context to help the user achieve their goal.

The above example mostly leverages intents to guide the flow; however, you can also guide the flow with entities and slots. For example, if the user gives you the information that they’re new to Rasa at the beginning, you may want to skip certain questions by storing this information in a slot.

Finally, to truly enhance your chatbot conversational abilities, consider implementing feedback loops to improve its learning capabilities over time.