We want a way for chat conversations to represent data that the user has entered via A2UI.
Scenario
-
Let's say I have a conversation and the agent asks me to pick a date range, showing "Start vacation date ": "end vacation date" , then a submit button.
-
I pick dates
-
I click submit, triggering an action with name "submit_dates" and the JSON of each date.
Typical behavior
In the conversation history, most agent implementations will show a message like "User responded" etc as a placeholder for the action that has been submitted, to give visual feedback that the agent has been prompted.
Good behavior
A message like "Selected a vacation period from x/x/xxx to y/y/yyyy" shows up.
Ideas
Perhaps the "action" schema could provide a property for providing a user-visible string? Perhaps that can reference functions, so that string format etc can be used to format a good string with values from the data model etc?
Though I wonder if the user visible string should actually be exactly the same as what the LLM sees? Perhaps that is the default way things work, then it's optional to add extra LLM-only context?
We want a way for chat conversations to represent data that the user has entered via A2UI.
Scenario
Let's say I have a conversation and the agent asks me to pick a date range, showing "Start vacation date ": "end vacation date" , then a submit button.
I pick dates
I click submit, triggering an action with name "submit_dates" and the JSON of each date.
Typical behavior
In the conversation history, most agent implementations will show a message like "User responded" etc as a placeholder for the action that has been submitted, to give visual feedback that the agent has been prompted.
Good behavior
A message like "Selected a vacation period from x/x/xxx to y/y/yyyy" shows up.
Ideas
Perhaps the "action" schema could provide a property for providing a user-visible string? Perhaps that can reference functions, so that string format etc can be used to format a good string with values from the data model etc?
Though I wonder if the user visible string should actually be exactly the same as what the LLM sees? Perhaps that is the default way things work, then it's optional to add extra LLM-only context?