Handling Business Logic

Conversational assistants often support user goals that involve collecting required information from the user before doing something for them. For example, a restaurant search bot would need to gather a few pieces of information about the user's preferences to find them a suitable restaurant:

This page is a guide on handling the business logic of collecting user information to fulfill a request. In the example above, business logic includes needing to know the user's preferred cuisine, party size, and seating preference. The examples on this page come from the formbot example bot.

Step-by-step Guide on Using Forms to Handle Business Logic#

Forms work by prompting the user for information until it has gathered all required information. The information is stored in slots. Once all the required slots are filled, the bot fulfills the user's original request.

1. Defining the form#

To define a form, you will need to define:

• Slot mappings: The required info to collect
• Responses: How your bot should ask for each piece of information

Slot Mappings#

For the restaurant search example, we want to collect the following information from the user:

• cuisine
• number of people
• whether they want to sit outside or not

You define a form in your domain by specifying a list of required slots under the form name:

These slots need to be added to your domain's slots section, along with slot mappings which define how the slots can be filled. For any slot filled from_entity, the entity also needs to be added to the domain. Slots filled by forms should usually not influence the conversation, so set influence_conversation to `false:

The number slot is filled from an entity. Entities like numbers can be extracted by DucklingEntityExtractor. To use it, add DucklingEntityExtractor to your NLU pipeline:

Slot Mappings with Conditions#

The outdoor_seating slot is filled based on the user's intent: If it is affirm, it'll be true, if it is deny, it'll be false.

However, the slot should only be set to true or false if the user was responding to the question, Do you want to sit outside?. To enforce this condition, the conditions for the outdoor_seating slot requires that restaurant_form is active and that the requested slot is outdoor_seating. If there were no conditions and the user had sent a message with the affirm or deny intent earlier in the conversation, the outdoor_seating slot would already be filled when the form was activated. Therefore the form would not prompt the user for their outdoor seating preference. See mapping conditions for more information.

Validating Slots#

Often, you'll want to validate the user's input before accepting it, for example by checking if the given cuisine is in your assistant's database of available cuisines. See the docs on validating form input for more information about validation actions.

Requesting Slots#

To specify how the bot should ask for the required information, you define responses called utter_ask_{slotname} in your domain:

2. Updating the configuration#

A form's happy path should be defined as a rule which means you'll need to add the RulePolicy to your policies:

3. Creating rules#

The form itself takes care of the logic around asking the user for all the required information, so you need only two rules for a form's happy path: One that defines when it starts, and one that defines what happens when it has been filled. For the restaurant search example, in real life the assistant would look up a restaurant based on the user's preferences. In this case, the bot will utter a response with the details that would be used for a search.

By splitting up the activation and submission of the form, the rules will still apply if the user provides unexpected input or interrupts the form with chitchat.

4. Updating the NLU training data#

You'll need to add examples for the intent that should activate the form, as well as examples for how the user will provide the required information.

Form Activation Intent(s)#

You need to provide training examples for the intent(s) that should activate the form. Add examples for the intent request_restaurant:

Slots filled with from_entity can by default be filled by any user utterance, regardless of the intent, as long as the correct entity is extracted. That means that if the user provides the cuisine entity as part of their first message, the slot will be filled at the beginning of the form and the bot will not ask them for the cuisine again.

Form Filling Intent(s)#

While the form is filling slots, it will not pay attention to which intent was predicted unless a slot mapping explicitly requires or excludes an intent.

For the restaurant search example, the outdoor_seating slot is mapped to two intents, so you need to add training data for these intents.

For the cuisine and number slots, no intent is specified, so you can add examples to a generic inform intent. You need to annotate the cuisine entity so that DIETClassifier can learn to extract it. You don't need to annotate the number entity since DucklingEntityExtractor is a rule-based extractors that isn't trained on your training data. Only a few examples are shown for each intent; for your bot to work well, you should add more training data than is shown here:

Update your domain to include these intents:

5. Defining the responses#

Add the responses that are sent after the form has been submitted:

Summary#

Forms can simplify the logic of collecting user information. To define a minimal form like the restaurant search example above, this is a summary of what you'll need to do:

• Add the RulePolicy to config.yml
• Define the form with required slots in the domain
• Add slot mappings for all required slots in the domain
• Add rules for activating and submitting the form
• Add examples for the intent(s) to activate your form
• Add examples for the intent(s) to fill the required slots
• Define an action or response for the bot to take when the form is completed
• Update your domain with new intents and actions you've defined

To try out your newly defined form, retrain the bot's model by running rasa train and start rasa shell. Because the DucklingEntityExtractor is being used to extract entities, you'll need to start Duckling in the background as well (see the instructions for running Duckling).