System Guardrail Trigger Log Report
System Guardrail & AI Agent Filter Trigger Report
This report will provide insights into any inputs and outputs from interactions with virtual agents that trigger a system guardrail.
Running the Report
To access this report, navigate to Analytics > Baseline CSV Reports > Create Reports.
Select the System Guardrail Trigger Log report from the list and specify your filters to run the report. The available filters include:
Start & End Dates - This is a required set of filters to scope the report to conversations created within the specified time frame.
The report is also available via our System Guardrail Trigger Log API endpoint.
Some Questions the Report Answers
- How often are people interacting with my virtual agents in a harmful or malicious way?
- What kinds of topics/phrases are users engaging in that are triggering the system security guardrails?
- What are the categories that are triggered the most?
- Which interfaces are seeing the most guardrail triggers?
Data Columns
Below is the list of columns you can expect to receive from this report.
| Column Name | Value Description |
|---|---|
| created_at_utc | The UTC timestamp of guardrail trigger event. |
| interface_name | The name of the interface (bot/channel/voice line) where the inquiry was submitted. |
| audio_log_url | A URL to a page allowing you to download an audio recording file. (Voice only.) |
| categories_fired | The category(ies) that were flagged for the guardrail trigger event. |
| category_scores | The per category severity/confidence scores behind each fired category. |
| reasoning | The model's explanation for why the guardrail was triggered. |
| generation_outcome | What happened to the model's response after the guardrail evaluation, e.g. "model engaged", "model refused", "not attempted". |
| applied_tiers | The guardrail sensitivity tiers applied for the event, e.g. "strict", "minimal", etc. per category. |
| triggering_text | The raw input text that cause the guardrail to fire. |