Hello,
We’ve activated all guardrails for the rephrasing feature, including configuring the “restrict topics” guardrail to block specific subjects such as war, politics, and religion.
When we prompt the VA with questions related to these restricted topics, it correctly avoids them—either by triggering fallback dialog tasks or redirecting the conversation. However, in the debug window, we don’t see any indication that a guardrail was triggered, even though the behavior suggests it was.
According to the documentation, there should be a message in the debug console confirming that a guardrail was activated.
Can someone clarify:
Is this expected behavior?
Are the guardrails working correctly even without debug messages?
Are there any example implementations we can use to confirm proper guardrail activation and visibility?
Looking forward to your guidance.
Best regards,
Namrata Hinduja Geneva, Switzerland (Swiss)
Based on the latest platform updates, the behavior you are observing is likely expected. While the guardrail is functioning correctly by identifying and blocking restricted topics, the detailed activation logs for these GenAI features have been centralized in a more comprehensive analytics section, rather than the real-time Debug Console.
Here’s a breakdown to answer your questions directly:
Is this expected behavior?
Yes, this is likely the expected behavior in the current version of the “AI for Service” (formerly XO 11) platform. The real-time Debug Console that you see during testing is primarily designed to trace the conversational flow, including intent recognition, entity extraction, and dialog task execution.
While older or different types of logs might appear there, the specific, detailed logs for GenAI features like Guardrailsare now channeled to a dedicated analytics section for better monitoring and governance.
Are the guardrails working correctly even without debug messages?
Yes. The fact that your Virtual Assistant (VA) correctly avoids the restricted topics by triggering fallback dialogs is the strongest indicator that the guardrail is working as intended. The functional behavior of the bot is the primary confirmation. The logs serve as a secondary, more detailed audit trail.
How to Confirm Proper Guardrail Activation
To view the detailed logs and confirm that the “restrict topics” guardrail was activated, you need to look in the LLM and Gen AI Logs, which you’ve correctly located in your screenshot. This is the designated area for monitoring the interactions between the platform and any Large Language Models.
Based on your image, here is the precise way to find the confirmation you’re looking for:
- Navigate to Analytics > GenAI Analytics > LLM and Gen AI Logs.
- Ensure the date range filter at the top left is set to the time you ran your tests.
- Click on the “More Filters” button.
- From the dropdown menu, select the relevant filters to narrow down the results. To check if a user’s inputtriggered a guardrail, you should select “Guardrails - LLM Input”. To check if the bot’s own response was blocked, you would use “Guardrails - LLM Output”. For your specific use case of blocking user queries about war, politics, etc., the “Guardrails - LLM Input” filter is the most relevant.