In-depth tutorial available
An in-depth step-by-step tutorial is available here. We recommend starting there and use this documentation article as a reference.
Stories are great to model conversation flows but they lack essential features to implement strong and enforceable business logic.
A simple business logic might be to collect the necessary information to prepare an API call, ensure this information is correct and notify the user when it’s not. Advanced business logic can take the form of a decision tree where the next question to be asked depends of an arbitrary combination of previous responses.
Rasa forms and the
FormAction are the official response however they rely heavily on code and lack the simplicity and readability of stories.
Forms in Botfront expose all the advanced Rasa Forms capabilities in a simple and intuitive graphical user interface.
There are two mains steps in implementing forms:
- Creating and configuring the form
- Integrating the form in your stories.
You can create a form by clicking the + button next to a group, then choose Form:
You can change the name of the form in the sidebar.
From the Rasa stand point, a form is a collection of slots waiting to be filled. Once Rasa enters a form, it iterates through slots (and ask the corresponding questions) until all are filled with valid data.
Let’s create a simple form collecting email address, first name and last name: To add a slot, click + icon at the bottom of the node, choose the slot where the user response will be stored, and type the question the assistant will ask.
Note that when you create new slots from this interface they will be unfeaturized. If you need to fill a featurized slot (e.g. categorical, boolean,…), you must create it from the slots menu and choose the Or use an existing one option.
In the following video we create a boolean allergies slot and use it in our form.
Slots created from the forms interface are also destroyed when you delete the question. However, pre-existing slots are not deleted when a question in removed.
By default, the entire user response is stored in the slot. You might need to change this behaviour to allow a more naturally flowing conversation.
When asking for a name, a user might answer “My name is Nathan”, where Nathan can be an extracted entity.
Click on the question to open the slot panel, and in the Extraction tab, choose Get the slot from the entity and select the entity you want to map to this slot.
In the example below, instead of filling the slot with the whole user response, we just extract the email entity. Because Duckling follows a pattern matching approach, well formed email addresses are guaranteed to be extracted. See here how to configure your NLU pipeline with Duckling.
By doing this, you ensure that your assistant will store a valid email address when users respond something like “my email address is email@example.com”
Use Duckling to extract structured entities such as numbers, dates, email addresses, urls, amounts of money and many more.
The last slot we added to our form was boolean. Boolean slots can only have
When asking for known alergies, users will most likely respond by yes or no, or more generally with affirm and deny intents.
Click on the question to open the slot panel, and in the Extraction tab, choose Get the slot conditionnally on the intent.
In the following video we show how to map affirm and deny intents to boolean values:
Combining extraction methods will allow your assistant to extract slot values with more robustness. Here is an example: when asking a first name, a user might answer John, or I’m John, or my name is John.
However such a name entity is hard to pick-up consistently when users only answer with their names (without my name is …)
To handle all situations, you can combine conditions as follows: the first one will check if there is an entity to fill the slot from. If it fails, it will take the entire response.
inform intent is found with a
name entity the slot will be filled from the entity. Otherwise, the whole message will be taken.
You may condition the extraction on the intent found:
When collecting the email from an entity, you may want to ensure that the intent corresponds to declaring an email address (e.g my email is firstname.lastname@example.org) and not something else such as I wrote an email to email@example.com and no one replied.
The inclusive condition below will make sure that the slot is filled from the
name entity if and only if the intent
inform is recognized.
If any other intent is found the form will be interrupted and you will have to handle that in your story.
The exclusive condition below will always fill the
message slot from the whole content of the user message, unless one of the listed intent is found.
It will then interrupt the form and you can take the conversation from there in your story.
This is particularly useful to allow users to stop filling forms.
To validate user input before storing it and going the next question, go to the Validation tab, enable validation and specify validation criteria.
You must also provide an message that the bot will utter when a response is invalid.
You can also specify a response to be uttered when the response is valid. You can refer to the slot value using the curly brace syntax (
Validation will only trigger after a slot is succesfully extracted. If extraction fails (e.g. the slot must be extracted from an entity that is not found) the form not try to validate user input. It will stop and you will need a story to handle it.
A variety of validation options are offered, and if they are not sufficient you can also match a regular expression.
Until now we had an linear flow of questions, but you can create much more complex flows with the flow chart editor.
Let’s consider a simple medical triage form asking for symptoms. We are going to ask:
- If the user has fever
- If the user has cough
- If the user has nausea
- A date preference for the appointment
In the first video we create the first two slots. See how we use conditional intent extraction to handle yes/no responses.
Let’s suppose that you only want to ask about nausea if the user present no cough symptoms.
But whether or not you ask about nausea, you want to move forward with the next question about appointment preference.
Note how we’re using the
NOT clause together with
fever being true.
Note how we used the edge edition mode to add a new edge from nausea to appointment_pref. You can also use it to re-order questions/slots: delete existing edges and replace them with new ones.
Finally, let’s add a last condition: we only want to ask for the time preference if the user has at least one symptom.
We can do this by adding conjunctions (
OR) to ALL edges going to appointment_pref.
Depending on how your form played out, maybe you also want to set a slot at the end of a particular path. This will allow your bot to know how the form ended. More details over here.
You can only set categorical or boolean slots using this feature. Click Set a slot and choose the slot you want to set and the value you want to set it to.
You can edit the slot and the value after placing it by clicking on the node.
Now that we have a form we may start using it in our stories.
The simplest integration covers the ideal situation where the user is collaborative and answers all questions. You only need to start the form in your story. Once it is active it is going to iterate through all slots, ask questions, validate answers, etc.
You can pick up after completion to continue the conversation once all the slots have been filled succesfully.
A key point to understand to debug your projects is when the flow stops to be handled by the form and stories are taking over.
A form is completed when there is no more slots left to fill. In the triage example above, if a user answers no to all symptoms, the questions about nausea and appointment_pref will not be asked because of the conditions set and the form will be completed after two questions (fever and caugh)
This can happen if you extract from an entity or conditionnally on an intent, and/or if you add intent conditions.
Suppose you extract an email address with an email entity, and the NLU does not find this entity. The extraction will fail and you will need a story handling this particular intent (or intente-entities combination) following your forms.
You will see how to do that in the next section.
To react to a particular exit from a form, you can set a slot at one of the ends of your form.
Look at this form, we want to follow up with another story if the user is not authorized by a legal guardian. We’re setting a slot to achieve this goal.
We can then use that slot in a story, the first story is used for the normal flow, when we end the form without setting anything, we just carry on with our happy path.
In the second story however, we can react to the forbidden slot being set, and continue with the forbidden path.
The recommended approach is to use branches and links to handle digressions. Please check our in-depth tutorial for an extensive example.
Contextual questions are questions with responses depending on information being collected. For example Explain or Why do you need this. Contextual questions must be explicitely enabled in the form configuration:
At training time, a categorical
requested_slot will be added to the domain with all slots contained in all forms as categories.
Once contextual questions are enabled/disabled for one form, it is enabled/disabled for all forms
Please check our in-depth tutorial for an extensive example.