Overview

Configures NeMo Guardrails rails to enforce safety and content moderation in conversations. It covers input, output, dialog, retrieval, and execution rails, plus Colang-based flows, bot templates, and user patterns for precise control. By enabling topic control, jailbreak detection, and fact-checking rails, it helps teams deploy compliant, safer AI assistants.

How This Skill Works

Define rail types (input, output, dialog, retrieval, execution) and Colang components (flows, templates, patterns, actions) to govern prompts and responses. Configure rails options including LLM selection, embedding model, and action handlers, then deploy within the target guardrails processes (system-prompt-guardrails and content-moderation-safety) to enforce policy during conversations.

When to Use It

• Deploying a customer-support bot with safety and factuality requirements
• Detecting jailbreak attempts and moderating content
• Designing Colang-driven flows and user/message patterns
• Integrating with a specific LLM, embedding model, and action handlers
• Testing rails with adversarial inputs and monitoring activations

Quick Start

1. Step 1: Install and import nemoguardrails and integrate the nemo-guardrails skill into your agent
2. Step 2: Define Colang flows and rails (input, output, dialog, retrieval, execution) using the skill's components
3. Step 3: Configure rails options (LLM, embedding model, action handlers) and test against adversarial prompts while monitoring activations

Best Practices

• Start with built-in rails
• Design clear flows
• Test with adversarial inputs
• Monitor rail activations
• Reuse Colang components to build maintainable rails

Example Use Cases

• A banking support bot that blocks sensitive prompts and flags high-risk input for manual review
• A healthcare assistant that filters medical advice and fact-checks before responding
• An e-commerce assistant that stays on product topics and avoids jailbreaking prompts
• An internal policy bot that enforces enterprise guidelines and logs guardrail activations
• A research assistant that retrieves sources and validates facts against trusted databases

Frequently Asked Questions