XO GPT Feedback Submission

Back to XO GPT Model Specifications

Use this guide to provide effective feedback on XO GPT models and understand how Kore.ai incorporates it.

---

How to Provide Effective Feedback

1. Assess Frequency

• Measure how often issues occur across a wide range of samples.
• Prioritize frequent issues—they have the most impact on model performance.
• Use occurrence rates to decide whether an issue needs immediate attention or further monitoring.

2. Identify Recurring Patterns

• Focus on consistent issues, not isolated errors (unless the issue is critical).
• Document patterns for a more accurate evaluation of model behavior.

3. Categorize the Issue

Category	Description
Misinterpretation	Model misunderstands the intent or meaning of the input.
Negation	Model incorrectly processes negation terms (e.g., “not,” “never”).
Omission	Model fails to include critical information in the response.
Redundancy	Model provides excessive or unnecessary information.
Repetition	Model repeats phrases or ideas unnecessarily.

4. Identify the Use Case

Specify the context—for example: Address Update, Medical Claim Processing, Customer Support Queries, or Financial Transactions.

5. Lock Sample Sets

• Once a problematic sample set is identified, use it consistently for testing.
• Track new issues separately with fresh sample sets to avoid overlap.

6. Submit a Support Ticket

When recurring issues are identified, submit a support ticket with the sample set, identified patterns, and the following information:

LLM Input	Actual Response	Expected Response	Error Category	Use Case	Comments
”Update my address to 1234 Elm St"	"Your address has not changed"	"Your address has been updated to 1234 Elm St”	Negation	Address Update	Model failed to process negation correctly.
”Process my medical claim for surgery"	"Your claim has been denied"	"Your claim is under review”	Misinterpretation	Medical Claim	Incorrect understanding of claim status.

When sharing data, mask all sensitive information. Do not include real customer data—it is used for training purposes. Kore.ai strictly prohibits the use of actual customer data for model training.

All processes and outputs adhere to company-wide standards. Changes beyond these guidelines may not be possible. Contact the support team for questions about data handling.

---

Feedback Workflow

Step	Action
1. Submit Feedback	Provide detailed feedback with examples, context, and expected output.
2. Analysis	Kore.ai reviews the feedback and defines the scope of improvements.
3. Data Collection	Additional data may be requested to improve model accuracy.
4. Model Update	The model is retrained and refined based on feedback and new data.

---

Continuous Improvement

Each feedback round should focus on new issues and use updated sample sets for accurate evaluation. Ongoing feedback ensures the model stays aligned with your evolving business needs.