Skip to main content
The Thresholds feature provides configurable score boundaries that determine how the NLU model interprets and classifies user inputs based on their Intent. This feature enables precise control over Intent recognition, ensuring the NLU can confidently identify Intents, prompt for reconfirmation when needed, or reject unclear inputs.

Key Benefits

  • Improved Accuracy. Adjustable thresholds enhance the reliability of Intent mapping by setting clear confidence levels.
  • Flexible Control. You can fine-tune the system’s sensitivity to match specific use cases or performance needs.

Prerequisites

Before using thresholds, ensure the following:
  • An active Flow is set up within the Project.
  • Intents are defined to allow the NLU to classify user inputs.

How to Use

Explore threshold types, then set them up.

Explore Threshold Types

Reconfirmation Threshold Specifics

The reconfirmation threshold behaves differently depending on whether a confirmation sentence exists in your Intent:
  • With a confirmation sentence, the reconfirmation threshold triggers the reconfirmation sentence.
  • Without a confirmation sentence, the reconfirmation threshold confirms the user input.

Set up Thresholds

Thresholds can be configured at either the Flow or Project level. To set them up:
  1. In the Flow editor, go to Settings > Configuration.
  2. Activate the Thresholds option. As soon as activated, this option overrides the Thresholds values set up at the Project level.
  3. Adjust the threshold sliders:
    • Reconfirmation Threshold (default: 0.2):
      • Move the slider left to increase reconfirmation prompts.
      • Move the slider right to reduce reconfirmation prompts.
    • Confidence Threshold (default: 0.4):
      • Move the slider left to allow lower-confidence matches.
      • Move the slider right to require higher confidence for matches.

Example

Imagine a user asks Can you book a flight? in the chat with the AI Agent. The NLU model assigns a score to the BookFlight Intent:

More Information

Last modified on April 21, 2026