I am working at creating a custom prompt for the Conversation Manager - DialogGPT feature. There is a section at the end about Essential Output Keys in Response Payload for AI for Service to Orchestrate Conversation Manger - DialogGPT Flow that requires values. Can anyone give me some insite in what needs to go itno these fields? I cannot find any documentation or examples of what these fields do.
Hi Barb,
Those “Essential Output Keys in Response Payload” are the mandatory JSON fields that your DialogGPT model must return so the Conversation Manager can correctly orchestrate the conversation. These keys tell the platform what kind of request it is, what intent was detected, and what to do next.
Here’s what each field means:
1. category
Defines which type of logic path the system should take next.
- Category 1: User intent matches a Dialog, FAQ, or Knowledge task.
- Category 2: System or fallback behavior, for example, “Continue” or “NoIntent_Identified.”
- Category 3: General or small talk intent, for example, “Live Agent Transfer” or “Greeting.”
2. fulfillment_type
Describes how the bot should fulfill the user’s request.
Examples include,
single_intent, multiple_intents, ambiguous_intents, generate_answer, system_intent, or conversation_intent.
3. winning_intents
Lists the intents that DGPT matched. The format is strict so the Conversation Manager can find the right dialog,
- Dialog intents → “appName:intentName”
- FAQ intents → “appName:Primary Question”
- Knowledge intents → “DocID-ChunkID”
- System or conversation intents → “IntentName” (no prefix)
Example:
“winning_intents”: [“TravelBot:book_flight”]
4. execution_plan
Defines the order of actions when multiple intents are detected.
Example:
“execution_plan”: [“TravelBot:book_flight”, “TravelBot:add_luggage”]
Sample Output
{
“category”: “Category 1”,
“fulfillment_type”: “single_intent”,
“winning_intents”: [“TravelBot:book_flight”],
“execution_plan”: [“TravelBot:book_flight”]
}
These fields are essential, if any are missing or empty, the Conversation Manager cannot continue the flow. Returning this structure ensures DGPT responses are parsed and executed correctly.
Very good explanation except for I don’t see anything about intent phrases.
Hi Barb, let me know if this does not make sense to you and I can go into more specific detail.
What are intent phrases, and why do we even care?
Imagine you’re talking to a virtual assistant at a restaurant. You say, “Cancel my salad, and order fries.”
To you, it’s one thought. To the system, it’s two actions:
• Cancel a food item → “cancel my salad”
• Order a new food item → “order fries”
Those snippets, “cancel my salad” and “order fries,” are the intent phrases. They show exactly which part of the user’s message triggered each intent.
Why it matters:
When the system passes control to the next step, like a Dialog Task or Agent Node, those intent phrases tell it what to cancel or what to add without needing to ask again.
Instead of saying, “Which item did you want to cancel?”, it already knows, the salad.
That’s the real power of intent phrases. They carry the specific language the user used straight into the fulfillment step. It’s not just accuracy, it’s memory with purpose, the system keeps the context alive, so it doesn’t sound like it has amnesia between steps.
