Great Wave AI Platform
  • Great Wave AI Platform
    • Introduction
    • High Level Overview
  • Great Wave AI Studio
    • Quick Start
    • Global Navigation
    • Agent Domain
      • Agent Interface
      • Agent Approaches
      • Multi Agent Architectures
      • Agent Modules
        • Instruct
        • Design
        • Guardrules
        • Evaluate
          • Evaluation Pagination and Filter
          • User Feedback
        • Integrate
        • API (Agent)
      • Agent Tag and Filter
      • Agent Show/Hide
    • Knowledge Domain
      • Knowledge Interface
      • Organizing Knowledge
      • Knowledge Modules
        • Documents
        • Index
        • API (Documents)
    • Buddy Bot
    • User Account Management
  • Technical FAQs
Powered by GitBook
On this page
  1. Great Wave AI Studio
  2. Agent Domain
  3. Agent Modules

Guardrules

The Guardrules module in the Great Wave AI Studio is designed to provide robust protection for your AI systems using two main types of guardrules: input guardrulesand output guardrules. Here’s how to configure and use these features to ensure the security and integrity of your AI operations.

Input Guardrails

  • Purpose: Input guardrails are designed to protect the system from potentially harmful or unwanted inputs, or can be utilized topically to ensure the user provides certain information. They serve as a first line of defense against attacks or inappropriate data entering your AI environment.

Output Guardrails

  • Purpose: Output guardrails are focused on managing the risks associated with the outputs generated by the Large Language Models (LLM) themselves. This is particularly important for mitigating reputational risks or preventing the generation of inappropriate or harmful content.

  • Setting Guardrails:

    1. Navigate to the Guardrule Screen: Access the Guardrule settings in the Great Wave AI Studio.

    2. Specify Your Rules: Add rules you want your agent to follow by clicking "Add". This could include offensive language, keeping on topic, talking about competitors etc.

    3. Guardrules Response Message: There are two options for the response message, Detection and Prevention Mode. In Prevention Mode you can set a custom response message that will be triggered whenever a rule is triggered. For example this message could inform the user that their input has been modified or rejected due to security policies. In Detection Mode you can allow the Agent being configured to answer using natural language, and the pass/fail nature of the rule will be logged in Evaluation, and if operating within a multi agent architecture the chain will stop at the point of reaching a Fail.

    4. Refresh Agent: After making changes, it is crucial to hit the ‘Refresh Agent’ button to ensure that the new settings are applied effectively.

    5. Review: In the Evaluation Module Guardrules will display a "-" if none are set, a "Pass" if Guardrules are configured and have passed or a "Fail" if Guardrules are configured and have failed. If a Fail is triggered within a chain of Agents this will "break the chain" and stop subsequent calls to Agents further along in the chain. To find out what a rule was triggered, click into the Interaction in the Evaluation Mode and see "Guardrules", this will tell you why the AI decided for or against your rules.

    6. Remove Rules: To remove rules just click the Red X next to the rule you want to remove.

Finalizing Settings

  • Apply and Test: Once you’ve set both input and output guardrails, use the test functionality within the studio to ensure that everything operates as expected. Adjust the settings as needed based on the outcomes.

  • Continuous Monitoring and Adjustment: Regularly review and adjust the guardrails to stay aligned with evolving content standards and security requirements.

By effectively configuring and managing both input and output guardrails on the Great Wave AI Studio’s Security screen, you can enhance the safety and reliability of your AI applications, ensuring they operate within desired ethical and legal boundaries.

PreviousDesignNextEvaluate

Last updated 11 days ago