How to beginner · 3 min read

How to add safety rails with NeMo

Q: How to add safety rails with NeMo

Use NeMo Guardrails to add safety rails by defining rules and constraints that filter or modify AI outputs. Integrate nemo_guardrails in your Python app to enforce safety policies on LLM responses.

Quick answer

Use NeMo Guardrails to add safety rails by defining rules and constraints that filter or modify AI outputs. Integrate nemo_guardrails in your Python app to enforce safety policies on LLM responses.

PREREQUISITES

Python 3.8+
pip install nemo_guardrails
An AI model client (e.g., OpenAI API key)
Basic knowledge of Python async programming

Setup

Install the nemo_guardrails package and set your AI API key as an environment variable. NeMo Guardrails works with OpenAI or other LLM clients.

bash

pip install nemo_guardrails

Step by step

Create a guardrails config file defining rules for safe AI responses, then load and run the guardrails in your Python app to filter outputs.

python

import os
import asyncio
from nemo_guardrails import Guard
from openai import OpenAI

# Set your OpenAI API key in environment variable OPENAI_API_KEY

# Define a simple guardrails config inline (usually in a YAML file)
guardrails_config = {
    "rules": [
        {
            "name": "no_hate_speech",
            "type": "block",
            "patterns": ["hate", "violence", "discrimination"]
        }
    ]
}

async def main():
    # Initialize OpenAI client
    client = OpenAI(api_key=os.environ["OPENAI_API_KEY"])

    # Initialize Guard with config and OpenAI client
    guard = Guard.from_dict(guardrails_config, client=client)

    # Input prompt
    prompt = "Write a sentence about peace and kindness."

    # Run guardrails to get safe response
    response = await guard.generate(prompt)

    print("Guardrails output:", response)

if __name__ == "__main__":
    asyncio.run(main())

output

Guardrails output: Peace and kindness bring people together and create harmony.

Common variations

Use YAML files for complex guardrails rules instead of inline dicts.
Integrate with other LLM providers by passing their client to Guard.
Run guardrails synchronously with asyncio.run() or in async frameworks.

Troubleshooting

If guardrails block all outputs, check your rule patterns for overblocking.
Ensure your AI client API key is set correctly in os.environ.
Use logging in nemo_guardrails to debug rule matching.

✅

Key Takeaways

Use nemo_guardrails to define and enforce safety rules on AI outputs.
Guardrails can block or modify unsafe content based on customizable patterns.
Integrate guardrails with any LLM client by passing it to the Guard class.
Use async Python to run guardrails smoothly in your applications.

Verified 2026-04 · gpt-4o, gpt-4o-mini

Verify ↗