📋 Forms

Forms are Particular Tools useful for collecting user information during a conversation!

How the Forms work

Imagine a scenario where you need to create an Order system for a pizzeria, using only the conversation with the user. The user must provide three pieces of information:

1. Type of pizza: must be a string from a predefined set.
2. Phone number: must be 10 digits long and follow a specific dialing code.
3. Address: must be a valid address in "Milano".

How can we solve this problem? The required information is very specific and needs:```

• Validators: We need validators to ensure the data is correct (e.g., phone numbers can vary by country, the pizzeria has a specific menu of pizzas, and delivery is restricted to certain areas of the city).
• Flexible Sequence: The information can be provided in any order during the conversation (e.g., a user might give the address before mentioning the type of pizza).

This is where Forms come in handy!

Implementation

import requests
from pydantic import BaseModel
from cat.experimental.form import CatForm, CatFormState, form

class PizzaOrder(BaseModel): #(1)
    pizza_type: str
    phone: str
    address: str


@form #(2)
class PizzaForm(CatForm): #(3)
    description = "Pizza Order" #(4)
    model_class = PizzaOrder #(5)
    start_examples = [ #(6)
        "order a pizza!",
        "I want pizza"
    ]
    stop_examples = [ #(7)
        "stop pizza order",
        "not hungry anymore",
    ]
    ask_confirm = True #(8)

    def submit(self, form_data): #(9)

        # Fake API call to order the pizza
        response = requests.post(
            "https://fakecallpizza/order",
            json={
                "pizza_type": form_data["pizza_type"],
                "phone": form_data["phone"],
                "address": form_data["address"]
            }
        )
        response.raise_for_status()

        time = response.json()["estimated_time"]

        # Return a message to the conversation with the order details and estimated time
        return {
            "output": f"Pizza order on its way: {form_data}. Estimated time: {time}"
        }

1. Pydantic class representing the information you need to retrieve.
2. Every class decorated with @forms is a Form.
3. Every Form must inherit from CatForm.
4. Description of the Form.
5. Pydantic class name.
6. Each Form must include a list of start examples to guide the LLM in identifying and initiating the form. This is close to the tool's docstring principle.
7. Each Form must include a list of stop examples to help the LLM in determining when to stop the form during the conversation.
8. A Form can request the user to confirm the provided data.
9. Every Form must override this method to define its functionality, such as calling a database to collect information, using an Order API, interacting with another agent or LLM, etc.

Changing the "actions" of the Form

Forms are implemented as FSM and you can modify any transition of the FSM by overriding the methods.

Here the diagram of the FSM: TODO

State-transition function

Each FSM has a State-Transition function that describes what is the next action to perform based on the given input. In the case of Cat's form implementation, the input is the User prompt and the def next(self) method acts as the State-Transition function.

The form evaluates four states:

1. INCOMPLETE
2. WAIT_CONFIRM
3. CLOSED
4. COMPLETE

Each state executes one or more phases:

• User Stop Form Phase
• User Confirmation Phase
• Updating Phase
• Visualization Phase
• Submit Phase

You can modify this state-transition by overriding the def next(self) method and accessing the state via self._state. The states are values from the CatFormState enum.

User Stop Form Phase

The User Stop Form Phase is when the Form checks whether the user wants to exit the form. You can modify this phase by overriding the def check_exit_intent(self) method.

User Confirmation Phase

The User Confirmation Phase is when the Form asks the user to confirm the provided information, if ask_confirm is set to true. You can modify this phase by overriding the def confirm(self) method.

Updating Phase

The Updating Phase is when the Form performs the Extraction Phase, Sanitization Phase and Validation Phase. You can modify this phase by overriding the def update(self) method.

Extraction Phase

The Extraction Phase is when the Form extracts all possible information from the user's prompt. You can modify this phase by overriding the def extract(self) method.

Sanitization Phase

The Sanitization Phase is when the information is sanitized to remove unwanted values (null, None, '', ' ', etc...). You can modify this phase by overriding the def sanitize(self, model)method.

Validation Phase

The Validation Phase is when the Form attempts to construct the model, allowing Pydantic to use the implemented validators and check each field. You can modify this phase by overriding the def validate(self, model) method.

Visualization Phase

The Visualization Phase is when the Form shows the model's status to the user by displaying a message.

By default the cat displays the forms like so

When there is invalid info retrieved from the conversation, the Cat specifies the issue

You can modify this phase by overriding the def message(self) method:

    # In the form you define 
    def message(self): #(1) 
        if self._state == CatFormState.CLOSED: #(2)
            return {
                "output": f"Form {type(self).__name__} closed"
            }
        missing_fields: List[str] = self._missing_fields #(3)
        errors: List[str] = self._errors #(4)
        out: str = f"""
        The missing information is: {missing_fields}.
        These are the invalid ones: {errors}
        """
        if self._state == CatFormState.WAIT_CONFIRM:
            out += "\n --> Confirm? Yes or no?"

        return {
            "output": out
        }

1. This method is useful for changing the Form visualization.
2. Forms have states that can be checked.
3. Forms can access the list of missing fields.
4. Forms can access the list of invalid fields and their associated errors.

Final Phase: Submit

The Submit Phase is when the Form concludes the process by executing all defined instructions with the information gathered from the user's conversation. The method has two parameters:

• self: Provides access to information about the form and the StrayCat instance.
• form_data: The defined Pydantic model formatted as a Python dictionary.

The method must return a dictionary where the value of the output key is a string that will be displayed in the chat.

If you need to use the Form in future conversations, you can retrieve the active form from the working memory by accessing the active_form key.

Here is an example:

    @hook  
    def before_cat_sends_message(message, cat):
        active_form = cat.working_memory.active_form