Implementing Guardrails
Ensure your AI agent stays helpful, safe, and on-topic using Guardrails.Why Guardrails?
LLMs are powerful but can sometimes “hallucinate” or be coaxed into discussing inappropriate topics. Guardrails act as a protective layer that monitors the conversation in real-time.Types of Guardrails
1. Topic Restrictions
Prevent the agent from discussing specific subjects (e.g., “Politics,” “Competitors,” or “Spiritual Advice”).- How to Config: Add a “Topic Restriction” in the Features tab and list the forbidden subjects.
2. Sentiment Guard
If the customer becomes extremely hostile, the agent can be instructed to automatically transfer the call to a human manager or politely terminate the interaction.3. Sensitive Data Redaction
Automatically detect and redact Social Security Numbers, Credit Card details, or Passwords from the call transcriptions for PCI/HIPAA compliance.4. Hallucination Check
Cross-references the agent’s output with your Knowledge Base. If the agent attempts to answer a question it doesn’t have data for, the guardrail can force a “I’m not sure about that, let me find out” response.Setting Up Guardrails
- Open your agent in the Agent Studio.
- Navigate to the Features > Guardrails section.
- Toggle on the specific protections you need.
- Optionally, specify a “Failure Message” the agent should say if a guardrail is triggered.