Nemo Guardrails

The nemo-guardrails skill provides a framework for implementing safety and policy enforcement mechanisms within AI agent workflows. It leverages the Nemo Guardrails library to define, validate, and enforce constraints on agent inputs and outputs dynamically.

When to use it

• Enforce strict content policies to prevent an agent from generating harmful, biased, or inappropriate text.
• Validate user inputs against specific schemas or allowed topics before processing them by the model.
• Create a "guardrail" layer that intercepts agent responses to ensure compliance with organizational standards.
• Test and iterate on safety rules without modifying the underlying language model logic.

Key capabilities

• Dynamic input validation to block unsafe queries before they reach the model.
• Output filtering to sanitize responses and remove prohibited content patterns.
• Custom rule definition allowing users to script specific safety constraints.
• Integration with orchestration frameworks to manage agent behavior securely.

Example prompts

• "Set up a guardrail that blocks any response containing personal identifiable information (PII)."
• "Configure an input validator to reject questions about sensitive medical advice."
• "Create a rule set that ensures all generated code snippets are wrapped in markdown code blocks."

Tips & gotchas

• Guardrails add latency; ensure your application architecture accounts for the extra processing time during validation steps.
• Rules must be carefully tuned to avoid false positives that might block legitimate user queries or useful agent responses.