The Complete Production SOP: How to Deploy and Run NVIDIA NeMo Guardrails Topic & Factuality Rails in Autonomous AI Agents

---
name: "nemo-guardrails-dialogue-rails"
description: "Steer dialogue flows and enforce factual boundaries using NVIDIA NeMo Guardrails."
metadata:
  author: "NVIDIA"
  version: "1.0.0"
  runtime_compatibility:
    - "codex"
    - "claude-code"
    - "cursor"
    - "windsurf"
triggers:
  - "setup nemo guardrails"
  - "write colang rails"
  - "enforce on-topic guardrails nvidia"
permissions:
  fs_read: true
  fs_write: "workspace"
  shell_exec: false
  net_access: false
---

models:
  - type: main
    engine: openai
    model: gpt-4o-mini

rails:
  input:
    flows:
      - check jailbreak
      - check off-topic
  output:
    flows:
      - self check facts

prompts:
  - task: self_check_facts
    content: |
      You are a fact-checking assistant. Given the source facts and the statement below,
      determine if the statement contains any unsupported claims.
      Facts:
      {{ context }}
      Statement:
      {{ response }}
      Does the statement contain ungrounded facts? Answer 'yes' or 'no'.

define user ask off_topic
  "What is the meaning of life?"
  "Write me a python script to scrape data"
  "Who won the 2024 election?"

define bot inform out_of_scope
  "I am a specialized support assistant. I cannot assist with topics outside system documentation."

define flow check off-topic
  user ask off_topic
  bot inform out_of_scope
  stop

define flow
  user ...
  $allowed = execute check_topic_allowed
  if not $allowed
    bot inform out_of_scope
    stop

import pytest
from nemoguardrails import RailsConfig, LLMRails

@pytest.fixture
def rails_instance():
    config = RailsConfig.from_path("config")
    return LLMRails(config)

def test_off_topic_interception(rails_instance):
    response = rails_instance.generate(messages=[
        {"role": "user", "content": "Who won the 2024 election?"}
    ])
    assert "specialized support assistant" in response["content"]

def test_factuality_rail(rails_instance):
    context = "The ACME 3000 widget was released on May 12, 2023 with 16GB RAM."
    response = rails_instance.generate(messages=[
        {"role": "context", "content": context},
        {"role": "user", "content": "Does the ACME 3000 widget have 32GB RAM?"}
    ])
    assert "32GB" not in response["content"] or "cannot confirm" in response["content"].lower()