DialogGPT is an agentic orchestration engine that powers natural conversations at scale.
It autonomously manages intent identification, execution planning, and fulfillment across Dialog Tasks—without requiring training data.
It uses text embeddings and generative models to identify user intents from task names and descriptions.
DialogGPT is the default intent identification mode for new AI for Service apps.
Overview
DialogGPT handles the full conversation lifecycle in three phases:
- Chunk Shortlisting: Processes user input and conversation history to retrieve relevant chunks from a vector database
- Intent Identification: Analyzes retrieved chunks with an LLM to resolve intent, ambiguities, and execution order
- Flow Management: Triggers fulfillment (dialog tasks, FAQs, answer generation) and delivers contextually relevant responses
Key Features
| Feature | Description |
|---|
| Zero-shot intent detection | Detects intents using RAG and LLMs—no utterance training required |
| Multi-intent handling | Recognizes and executes multiple intents in a single utterance, managing dependencies and execution order |
| Ambiguity resolution | Prompts users with clarifying questions when intent is unclear |
| Conversational nuance management | Handles pauses, repetitions, restarts, and digressions naturally |
| Dynamic response generation | Grounds responses in user data, conversation history, and business rules |
| Model flexibility | Supports commercial, custom, and XO GPT models |
| Granular intent resolution | Refines broad queries into specific, actionable intents using domain knowledge graphs |
Key Benefits
| Benefit | Description |
|---|
| No training overhead | Only concise dialog descriptions and FAQ alternate questions are needed; no utterance training required |
| Higher accuracy | Handles complex, multi-intent, and industry-specific queries with full context awareness |
| Lower operational costs | Increases self-service rates, reducing transfers to human agents |
How DialogGPT Works
DialogGPT processes each user utterance through three sequential steps.
Step 1: Chunk Shortlisting
DialogGPT rephrases the user input to optimize retrieval, then uses a RAG pipeline to fetch relevant chunks—segments of dialog tasks, FAQs, or Search AI embeddings stored in a vector database. This retrieval operates independently of the Search AI pipeline.
From v11.15.1, only sub-intents relevant to the active dialog are appended to {{dialogs_chunks}}. Previously, all sub-intents were indexed with the top-level intent.
Step 2: Intent Identification
The Intelligent Conversation Orchestrator analyzes retrieved chunks with the user input and conversation context. Using an LLM, it:
- Identifies the most appropriate intent.
- Resolves ambiguities by ranking options or prompting the user for clarification.
- Determines dependencies and execution order for multi-intent scenarios.
- Routes system intents (repeat, pause, end) to pre-defined handlers.
Fulfillment is categorized as: single intent, multiple intents, ambiguous intents, or system intents.
Step 3: Flow Management and Fulfillment
DialogGPT triggers the resolved intent and:
- Executes dialog tasks or FAQs.
- Generates answers for system intents using Answer Generation.
- Requests clarification for ambiguous intents.
- Handles multi-intent scenarios sequentially or in parallel based on dependencies.
Responses adhere to enterprise business rules and the configured interaction mode (text or voice).
Configuration
Conversation Types
Configure which conversation types DialogGPT handles. Dialogs and FAQs are enabled by default and cannot be disabled.
| Conversation Type | Description |
|---|
| Dialogs | AI Agent building blocks for task execution and user interaction |
| FAQs | Pre-defined Q&A for common inquiries |
| Conversation Intents | Common conversational phrases (hold, repeat, end, agent transfer) |
| Knowledge from Search AI | Answers sourced from enterprise data via Search AI |
Model Configurations
Use the Model Configurations card to set up models for chunk shortlisting and conversation management.
Shortlisting Relevant Chunks: BGE-M3 is the supported and recommended embedding model. The embedding model settings apply only to Dialogs and FAQs; for Knowledge from Search AI, the settings in the Search AI app apply.
Optionally adjust Similarity Threshold and Proximity Threshold in advanced settings. Defaults work for most cases.
Maximum Chunks from Search AI: Set the maximum number of chunks shortlisted from Search AI and sent to the LLM for answer generation. Default: 5. Click Go to Search AI to configure retrieval settings.
Conversation Management: Select the LLM for intent detection and execution planning. Supported models:
| Model |
|---|
| OpenAI GPT-4o |
| OpenAI GPT-4o mini |
| Azure OpenAI GPT-4o |
| Amazon Bedrock |
| Custom LLM |
| XO GPT – DialogGPT |
Fulfillment
Fulfillment defines how DialogGPT responds to identified intents. There are two event types: Intent Events and Conversation Events.
Intent Events
| Event | Default Behavior | Override Options |
|---|
| Intent not identified | Trigger fallback task | — |
| Ambiguous intents | Present list of qualified intents for user selection | Execute a specific dialog task |
| Answer Generation | Activated when Knowledge from Search AI is enabled; added to the app automatically | — |
| Multiple intents | Execute predefined MultiIntent dialog | Modify the MultiIntent fulfillment dialog |
Conversation Events
Conversation Events activate when Conversation Intents is enabled. They handle common conversational phrases.
| Event | Default Behavior |
|---|
| Interaction Intents | Execute ConversationEvent dialog task (handles hold, context answers, refusing to answer) |
| Restart Conversation | Ask for confirmation → restart welcome flow |
| Agent Transfer | Ask for confirmation → execute DefaultAgentTransfer dialog |
| End Conversation | Ask for confirmation → end conversation |
Enabling DialogGPT
Prerequisites: Integrate the LLM that will power DialogGPT. See Model Configurations and LLM Integration.
Steps:
- Go to Generative AI Tools > DialogGPT.
- Click Get Started.
- Select the Conversation Types to enable.
- Under Model Configuration, select the embeddings model for Dialogs and FAQs.
- (Optional) Click Show Advanced Settings to adjust Similarity Threshold and Proximity Threshold.
- (If Knowledge from Search AI is selected) Set the maximum number of chunks to shortlist from Search AI. Click Go to Search AI for retrieval settings.
- Select the Conversation Management Model and Prompt.
- (Optional) Adjust Temperature, Max Tokens, and Conversation History Length.
- Click Enable DialogGPT.
To disable: Click the More icon (⋯) in the top right > Disable DialogGPT > confirm. Disabling reverts intent detection to traditional NLU.
Debug Logs
Debug Logs provide step-by-step visibility into DialogGPT’s decision-making and execution flow. Every LLM call includes full request and response details.
For each user utterance, logs capture:
| Log Entry | Description |
|---|
| Shortlisted chunks | Dialogs, FAQs, and Search AI chunks retrieved |
| Resolved chunks | Final chunks used for intent resolution |
| Fulfillment type | Intent, FAQ, or Answer |
| Exit path | Intent identified, FAQ match, or generated answer |
| Prompt payload | Full LLM prompt in JSON format |
| Fulfillment details | Execution outcome |
| Source information | Source reference (for Answer fulfillment type only) |
FAQs
Does DialogGPT require training data?
No. DialogGPT uses zero-shot intent detection via pre-trained embeddings and LLMs. Developers provide concise dialog descriptions and FAQ alternate questions—embedding models handle the rest.
How does zero-shot understanding work?
Pre-trained embeddings process the query and conversation history to identify intents and generate accurate responses for unseen scenarios, without any prior training examples.
How does DialogGPT handle multiple intents in one query?
It uses LLMs to parse multi-intent queries and executes intents in parallel where possible. For example, “Check the weather and book a flight” triggers both tasks simultaneously.
How does DialogGPT resolve ambiguous queries?
It ranks ambiguous options and either presents them to the user for selection or prompts for clarification before proceeding.
How does DialogGPT handle FAQs?
FAQs are treated as primary resources. User inputs are processed against FAQ chunks in parallel with dialog intents and search queries, using embeddings for retrieval.
How does DialogGPT handle conversational interruptions?
It supports conversation intents for common interruptions, including:
- Repeating information
- Holding or skipping tasks
- Handling digressions (e.g., off-topic questions mid-dialog)
What role does query rephrasing play?
DialogGPT rephrases queries to optimize chunk retrieval. For example, “What about Hyderabad?” is rephrased with conversational context before processing, improving retrieval accuracy.
Can DialogGPT be tuned for domain-specific use cases?
Yes. Customize dialog descriptions, FAQs, and embeddings to suit the domain. The modular architecture supports adaptation across industries.
Does DialogGPT support small talk?
Yes. Small talk is treated as a conversation intent, making interactions feel more natural and less robotic.
Can existing apps be migrated to DialogGPT?
Migration is supported. Transitioning from XO10 to XO11 is required; apps using Smart Assist may need additional configuration.
How does DialogGPT compare to traditional NLP?
| Aspect | Traditional NLP | DialogGPT |
|---|
| Training | Thousands of utterances per intent | Concise descriptions only |
| Multi-intent | Limited | Natively supported |
| Context handling | Rule-based | Embedding + LLM-based |
| Fallback | Predefined | Intelligent routing to search or fallback dialog |
| Orchestration | Complex to configure | Autonomous |
