Steps to Train a Question

Use this workflow to review, correct, and retrain Topics in the Training Panel. This process ensures AI responses stay accurate, explainable, and aligned with your instructions.

Steps to Train a Question

1. Select your topic and ask a question

Go to the Training Panel and select a Topic. In the Interactions tab, choose a real customer interaction to test or write your own question.

πŸ’‘ Pro Tip: Use real customer questions to get the most realistic training signal.

2. Check the topic assignment

Each question is automatically assigned to a topic. The topic determines which briefs, instructions, skills, and knowledge sources the AI uses when answering.

3. If the topic is incorrect: edit the topic description

If the assigned topic does not match the intent of the question:

Go to: Settings β†’ Topics β†’ Edit Topic
Update the topic description to clearly define the scope, intent, and types of questions it should cover.

Clear topic definitions ensure consistent answers across similar questions.

4. If the topic is correct: review the answer

Read the AI-generated answer carefully.

If the answer fully meets your expectations, give it a thumbs up.
If the answer is incorrect or incomplete: continue to the next step.

5. Inspect the Chain of Thought

Click Chain of Thought to understand how the AI reasoned through the question.

Use this view to review:

  • Sources used

  • Available skills

  • Enabled knowledge sources

  • Prior feedback that influenced the response

This helps you identify missing context, incorrect assumptions, or contradictions in the AI’s reasoning.

πŸ’‘ Pro Tip: We recommend toggling off Historical Tickets as a source and relying on Articles and Briefs instead. Historical tickets may include human errors, outdated decisions, or inconsistent handling, which can mislead the AI Agent.

6. Use Improve to refine briefs and instructions

Use Improve to get deeper insight into why Quack responded the way it did and to guide better future behavior.

With Improve, you can:

  • Ask Quack to explain gaps or uncertainties in the answer

  • Provide direct feedback on what the correct or expected answer should be

  • Help Quack resolve contradictions across knowledge, briefs, or instructions

πŸ’‘ Pro Tip: Use Improve to generate clearer, more precise briefs and instructions.

7. Start a new session to re-test and give feedback

Click New Session in the Training Panel and ask the same question again.

Only after the AI produces the expected answer should you provide positive feedback.
This ensures the AI is learning from the updated topic context, briefs, and instructions.

8. Train additional questions within the same topic

Continue testing different questions and use cases under the same topic.

Focus on:

  • Edge cases

  • Different customer scenarios

  • Variations in phrasing and intent

Repeat the steps above until the topic consistently produces accurate, reliable answers. At this point, the topic is production-read.