Hi,
I’m testing a demo To-Do app based on Chat Completion api with ‘gpt-3.5-turbo’ model.
The idea of the app is to manage the user’s tasks via conversation, and display them as todo list.
I usually test on a data set and reuse it again with different conversations after tuning my prompts.
In one round, I did the following conversation, and told the app, I’m done with following app with OpenAI. And the result was as expected:
In the next round (new conversation), I added the same tasks, but declared another task as completed, but I got both the previous conversation completed task and the current one as completed:
API is stateless, it has no memory of any other conversations or data, other than what you send in the request. Are you including other messages in the request?
Yes, there are default messages that are not shown to the user:
{ "role": "system", "content": "You are advanced To Do list assistant, you help in managing to do lists. Current date: " + (new Date()).toISOString().split('T')[0] },
{ "role": "user", "content": "I'll keep telling you about my tasks: what I need to do, doing, and/or done and you will provide me with a list of my tasks (pending and completed) as a to-do list" },
{ "role": "user", "content": "Each task must be shown in a separate line prefixed by [ ] for pending task and [*] for done task"},
{ "role": "user", "content": "Please add metadata about the task in the format [Metadata Name: value], the needed metadata are 'Creation Datetime', 'Completion Datetime', 'Due by'"},
{ "role": "user", "content": `Format your response as follows (the [{To-Do List Name}] line is needed when creating the to-do list, but not after modifying it):
[{To-Do List Name}]
----
{todo list}
----`}
