Let’s make a lengthy look at the internal dialog that might happen, since multiple internal iterations are possible.
Below is an expanded version of the initial exchange, detailing the intermediate steps where the assistant interacts with the scheduler functions you may have.
Following that, I’ve added a new user request to move the appointment an hour earlier, because the user changes their mind. Read through, and see the prior knowledge of tool use and tool return that would be needed:
- Does the AI know still the initial desired time to extrapolate “an hour earlier”?
- Is an appointment available, will the AI delete the original and make a new one? Can it “delete” if the information is lost?
- Does the AI need to know information about the success and appointment ID in order to make the change?
- Does the AI already have a listing of the day’s other appointments from a prior tool return?
Such an additional request is contextually-dependent. Not having enough information about tools, or the tools not showing usage, can make the AI model then think that it has the ability to schedule appointments without invoking a tool, or there may be loss of useful information, such as a previous document search with a troubleshooting database answer that may continue to be needed for reference.
It is also good to not force the AI to make repeated calls, as each multi-step tool invocation is paying for the chat history multiple times, with delays.
There is less need to manage the context window length to be short with new models, but budget and attention without distraction is a concern. You can look at how long the interactions in your application must have perfect understanding of the past, and answer this for yourself.
Additionally, context caching may offer a 50% discount on all past input that is identical. Deleting chat turns produces a different input sequence, breaking your discount when the chat history differs.
Initial Appointment Inquiry and Booking
Turn 1: User Inquiry
User:
Does Service A have availability at 5 PM?
Turn 2: Assistant Processes the Inquiry
Assistant:
Initiates a function call to check availability.
Message to functions.scheduler:
{
"action": "search",
"service": "Service A",
"datetime": "2024-04-27T17:00:00"
}
Turn 3: Scheduler Function Returns Availability
Tool (functions.scheduler) Response:
{
"available": true,
"service_id": "service_a_123",
"staff_available": ["staff_1", "staff_2"]
}
Turn 4: Assistant Responds to User
Assistant:
Yes, there’s availability for Service A at 5 PM.
Turn 5: User Requests Booking
User:
Great, please book it for me.
Turn 6: Assistant Processes the Booking
Assistant:
Initiates a function call to book the service.
Message to functions.scheduler:
{
"action": "book",
"service_id": "service_a_123",
"datetime": "2024-04-27T17:00:00",
"client_info": {
"name": "John Doe",
"contact": "john.doe@example.com"
},
"staff_assigned": "staff_1"
}
Turn 7: Scheduler Function Confirms Booking
Tool (functions.scheduler) Response:
{
"success": true,
"appointment_id": "appt_456",
"confirmation_message": "Your appointment for Service A at 5 PM has been booked successfully."
}
Turn 8: Assistant Confirms to User
Assistant:
Your appointment for Service A at 5 PM has been booked successfully. Your appointment ID is appt_456.
New Request: Moving Appointment an Hour Earlier
Turn 9: User Requests to Reschedule
User:
I’d like to move my appointment an hour earlier.
Turn 10: Assistant Processes the Rescheduling Request
Assistant:
Retrieves existing appointment details.
Message to functions.scheduler:
{
"action": "retrieve",
"appointment_id": "appt_456"
}
Turn 11: Scheduler Function Returns Current Appointment Details
Tool (functions.scheduler) Response:
{
"appointment_id": "appt_456",
"service_id": "service_a_123",
"datetime": "2024-04-27T17:00:00",
"client_info": {
"name": "John Doe",
"contact": "john.doe@example.com"
},
"staff_assigned": "staff_1"
}
Turn 12: Assistant Checks Availability for New Time
Assistant:
Calculates the new desired time (one hour earlier: 4 PM). Initiates a function call to check availability.
Message to functions.scheduler:
{
"action": "search",
"service": "Service A",
"datetime": "2024-04-27T16:00:00"
}
Turn 13: Scheduler Function Returns Availability for New Time
Tool (functions.scheduler) Response:
{
"available": true,
"service_id": "service_a_123",
"staff_available": ["staff_1", "staff_2"]
}
Turn 14: Assistant Initiates Rescheduling
Assistant:
Initiates a function call to update the appointment time.
Message to functions.scheduler:
{
"action": "update",
"appointment_id": "appt_456",
"new_datetime": "2024-04-27T16:00:00",
"staff_assigned": "staff_1"
}
Turn 15: Scheduler Function Confirms Rescheduling
Tool (functions.scheduler) Response:
{
"success": true,
"confirmation_message": "Your appointment has been moved to 4 PM."
}
Turn 16: Assistant Confirms Rescheduling to User
Assistant:
Your appointment has been moved to 4 PM successfully.
Summary of Tool Function Calls for Rescheduling
-
Retrieve Existing Appointment:
- Action:
retrieve
- Parameters:
appointment_id
-
Check Availability for New Time:
- Action:
search
- Parameters:
service, datetime
-
Update Appointment Time:
- Action:
update
- Parameters:
appointment_id, new_datetime, staff_assigned
This expanded flow demonstrates how the assistant interacts with a hypothetical scheduler functions to handle user inquiries, bookings, and rescheduling requests. It gives an “update appointment” method to change a time, but you can imagine that a booking service may work differently - like you can’t just change your Taylor Swift tickets to a different seat number, you’d have to cancel the existing order.