XO GPT DialogGPT Model - Kore.ai Docs

Back to XO GPT Model Specifications The XO GPT DialogGPT model predicts user intent in multi-turn conversations. For each user query, it takes shortlisted RAG chunks (Dialog, FAQ, and Knowledge), conversation history, and the active dialog context to determine the most relevant intent and generate an accurate response.

Challenges with Commercial Models

Challenge	Impact
Latency	High processing times affect user experience in real-time or high-volume scenarios.
Cost	Per-request pricing scales poorly for large deployments.
Data Governance	Sending queries to external models raises privacy and security concerns.
Lack of Customization	General-purpose models aren’t tuned for specific industries or use cases.
Limited Control	Minimal ability to correct or refine model behavior for incorrect outputs.
Compliance Constraints	Some industries have regulatory requirements that commercial LLM providers don’t fully support.

Key Assumptions

Scope: Supports text and voice-based conversations only.
All required pipeline inputs are available for the model to generate the expected output.

Benefits

Consistent and Accurate

Ensures accurate intent identification for every user input. See Model Benchmarks for metrics.

Cost-Effective

For Enterprise Tier customers, XO GPT eliminates commercial model usage costs. DialogGPT without Search (1,000 input tokens/utterance, 50,000 daily utterances, 10 output tokens/response):

Model	Input $/MTok	Output $/MTok	Input $/Year	Output $/Year	Total $/Year
GPT-4 Turbo	$30	$60	$547,500	$10,950	$558,450
GPT-4	$10	$30	$182,500	$5,475	$187,975
GPT-4o Mini	$0.15	$0.60	$2,738	$110	$2,847

DialogGPT with Search (5,000 input tokens/utterance):

Model	Input $/MTok	Output $/MTok	Input $/Year	Output $/Year	Total $/Year
GPT-4 Turbo	$30	$60	$2,737,500	$10,950	$2,748,450
GPT-4	$10	$30	$912,500	$5,475	$917,975
GPT-4o Mini	$0.15	$0.60	$13,688	$110	$13,797

Enhanced Security

No client or user data is used for model retraining. Guardrails: Content moderation, behavioral guidelines, response oversight, input validation, and usage controls. AI Safety: Ethical guidelines, bias monitoring, transparency, and continuous improvement.

Performance, features, and language support may vary by implementation. Test thoroughly in your environment before production use.

Use Cases

Domain	Use Cases
Healthcare	Appointment booking, symptom checker, retrieving upcoming appointment details, accessing medical records
Banking	Check balance, transfer funds, open or close accounts, and other retail banking tasks
HR	Request time off, manage leaves, check company policies, raise HR issues
Insurance	View policy details, modify insurance policies, file a claim, track claim updates

How It Works: Fulfillment Types

The model classifies each user query into a fulfillment type and category, then identifies the winning intent(s).

Fulfillment Type	Category	Description	Behavior
Single Intent	1	One intent identified from Dialog or FAQ Chunks.	The identified intent is directly executed.
Generate Answer	1	One intent identified from Knowledge Chunks.	The Answer Generation model synthesizes a response from the Knowledge Chunks.
Multiple Intents	1	Two or more intents identified from any chunk type; execution plan generated.	Intents are executed in the sequence specified by the execution plan.
Ambiguous Intents	1	Multiple possible intents with insufficient confidence to prioritize one.	All winning intents are presented to the user for confirmation before execution.
Continue	2	User query directly responds to an in-progress AI Agent inquiry.	Conversation continues seamlessly based on the input.
No Intent	2	User input doesn’t match any category.	A configured Fallback Intent is triggered.
Conversational Intents	3	Input matches a predefined conversational intent (see below).	Corresponding event is automatically triggered in the conversation flow.

Conversational intent types:

Intent	Triggered When
Answer from Context	User asks to repeat the last message or requests information from conversation history.
Refuse to Answer	User declines to provide requested information.
Pause Interaction	User requests a temporary halt in the conversation.
Restart Interaction	User asks to start the conversation over.
End Interaction	User indicates they’re done with the interaction.
Agent Transfer	User explicitly or implicitly requests escalation to a human agent.

Sample Outputs

Example 1: Ambiguous Intents

Conversation history:

[
  "AI Agent: Hello! How can I assist you today?",
  "User: I would like to schedule a virtual consultation with a healthcare provider.",
  "AI Agent: Sure! Could you please provide a preferred date and time?",
  "User: I am available on October 15th at 3 PM.",
  "AI Agent: You are booked for a virtual consultation on October 15th at 3 PM. Is there anything else?"
]

Dialog Chunks:

[
  {"ID": "kZryjJQBY6hZN8s7xxxx", "Name": "Virtual Consultation", "score": 0.5942571,
   "Description": "Schedule or initiate virtual consultations with healthcare providers"},
  {"ID": "jZryjJQBY6hZN8s7xxxx", "Name": "Get Definition of Disease", "score": 0.48599482,
   "Description": "Get descriptions of a medical condition on request"}
]

FAQ Chunks:

[
  {"ID": "lJryjJQBY6hZN8s7xxxx", "score": 0.63696957,
   "Primary Question": "How can I find a doctor who specializes in my condition?"}
]

User Query: I want to consult a doctor online. How do I do it? Output:

{
  "category": "Category 1",
  "fulfillment_type": "ambiguous_intents",
  "winning_intents": ["9Ham9fDB:Can I consult a doctor online?", "qxLm9qat:Virtual Consultation"]
}

Example 2: Multiple Intents

Conversation history:

[
  "AI Agent: Welcome to our banking service! How may I assist you today?",
  "User: Hi, I'm interested in some financial products.",
  "AI Agent: Certainly! What specific financial products are you interested in?"
]

Dialog Chunks:

[
  {"ID": "eksOZ5MB5L_rgqT0xxxx", "Name": "Apply Credit Card", "score": 0.67097855},
  {"ID": "e0sOZ5MB5L_rgqT0xxxx", "Name": "Apply Home Loan", "score": 0.64631605},
  {"ID": "eUsOZ5MB5L_rgqT0xxxx", "Name": "Apply Car Loan", "score": 0.55991244}
]

User Query: I want to apply for a credit card and a home loan. Output:

{
  "category": "Category 1",
  "fulfillment_type": "multiple_intents",
  "winning_intents": ["HvQ0049Q:Apply Home Loan", "RzWixwEa:Apply Credit Card"],
  "execution_plan": ["RzWixwEa:Apply Credit Card", "HvQ0049Q:Apply Home Loan"]
}

Example 3: Single Intent

Conversation history:

[
  "AI Agent: Hello! How can I assist you with your banking needs today?",
  "User: Hi, I'd like to set up automatic payments for my utility bills.",
  "AI Agent: Could you please provide the payment type and account?",
  "User: I'll be setting up payments for electricity and water. Use my checking account ending in 1234.",
  "AI Agent: How much would you like to pay for each bill, and how often?"
]

Active dialog context:

{'dialog_name': 'Set Payments', 'current_node': {'name': 'payment_amount_and_frequency', 'type': 'entity'}}

User Query: As of now, I need to update my mailing address. Output:

{
  "category": "Category 1",
  "fulfillment_type": "single_intent",
  "winning_intents": ["ooNsySJK:Change Address"]
}

Example 4: System Intent (Continue)

Conversation history:

[
  "AI Agent: Welcome to ABC Bank. How can I assist you today?",
  "User: Hi, I'd like to open a new account.",
  "AI Agent: We offer savings, current, and fixed deposit accounts. Which would you like?",
  "User: I would like to open a savings account.",
  "AI Agent: We need some personal information. Could you please provide your full name and contact details?",
  "User: Sure, my name is John Doe, and my contact number is 123-456-7890.",
  "AI Agent: We also need proof of ID and proof of address. Do you have these ready?",
  "User: Yes, I have my passport and a utility bill.",
  "AI Agent: Lastly, what initial deposit amount would you like to make?"
]

Active dialog context:

{'dialog_name': 'Open Account', 'current_node': {'name': 'initial_deposit_amount', 'type': 'entity'}}

User Query: The branch code is 00123. Output:

{
  "category": "Category 2",
  "fulfillment_type": "system_intent",
  "winning_intents": ["NoIntent_Identified"]
}

Example 5: Conversation Intent

Conversation history:

[
  "AI Agent: Hello! How can I assist you today in managing your HR tasks?",
  "User: change supervisor",
  "AI Agent: I understand you want to update the reporting structure. Could you please provide the employee's name?"
]

Active dialog context:

{'dialog_name': 'Supervisor Change', 'current_node': {'name': 'employee_name', 'type': 'entity'}}

User Query: I need to pause this for a moment. Output:

{
  "category": "Category 3",
  "fulfillment_type": "conversation_intent",
  "winning_intents": ["Pause_Interaction"]
}

When an AI Agent has PII masking enabled, masked input fields aren’t available for intent identification, which may affect the model’s accuracy.

Model Building Process

See Model Building Process.

Model Benchmarks

Version	Accuracy	TPS	Latency (s)	Benchmark	Test Data
v1.0	95.8%	35.1	1.57	Summary v1	Results v1

Version 1.0

Model Choice

Base model: Llama-3.1-8B-Instruct

Base Model	Developer	Language	Release Date	Status	Knowledge Cutoff
Llama-3.1-8B-Instruct	Meta	Multi-lingual	July 2024	Static	December 2023

Fine-Tuning Parameters

Parameter	Description	Value
Fine-Tuning Type	Method used	PEFT-QLoRA
Quantization	Bits for loading parameters	4-bit
Rank	Number of trainable parameters	32
LoRA Alpha	Scaling factor for LoRA updates	16
LoRA Dropout	Prevents co-adaptation in neural network	0.05
Learning Rate	Rate toward loss minimum	2e-4 (0.0002)
Batch Size	Examples per training step	1
Epochs	Passes over training data	5
Max Sequence Length	Maximum input length	32768
Optimizer	Optimization algorithm	paged_adamw_8bit
Task Type	LoRA task type	CAUSAL_LM

General Parameters

Infrastructure: 2× A10 GPUs.

Parameter	Description	Value
Learning Rate	Rate toward loss minimum	2e-4 (0.0002)
Batch Size	Examples per training step	1
Epochs	Passes over training data	5
Max Sequence Length	Maximum input length	32768
Optimizer	Optimization algorithm	paged_adamw_8bit

Benchmarks Summary v1

Comparison models: Phi4, GPT-4o, Llama 3.1 8B.

See Test Data and Results v1 for full details.

​Challenges with Commercial Models

​Key Assumptions

​Benefits

​Consistent and Accurate

​Cost-Effective

​Enhanced Security

​Use Cases

​How It Works: Fulfillment Types

​Sample Outputs

​Example 1: Ambiguous Intents

​Example 2: Multiple Intents

​Example 3: Single Intent

​Example 4: System Intent (Continue)

​Example 5: Conversation Intent

​Model Building Process

​Model Benchmarks

​Version 1.0

​Model Choice

​Fine-Tuning Parameters

​General Parameters

​Benchmarks Summary v1

Challenges with Commercial Models

Key Assumptions

Benefits

Consistent and Accurate

Cost-Effective

Enhanced Security

Use Cases

How It Works: Fulfillment Types

Sample Outputs

Example 1: Ambiguous Intents

Example 2: Multiple Intents

Example 3: Single Intent

Example 4: System Intent (Continue)

Example 5: Conversation Intent

Model Building Process

Model Benchmarks

Version 1.0

Model Choice

Fine-Tuning Parameters

General Parameters

Benchmarks Summary v1