Installation and setup

Installing NeMo Guardrails does not choose a model. The model is a few lines of config.yml, and a free Groq or Gemini key fits there.

pip install nemoguardrails 0.24.0

The package brings a runtime, a parser for the Colang configuration language and an embedding index, but no model client.

pip install "nemoguardrails==0.24.0"

Example

from importlib.metadata import version

print(version("nemoguardrails"))

A model entry in config.ymlOptional

This course registers a stand-in chat model in lessons 5 and 6 and a stand-in embedding model in lesson 10, so it runs without a key. For a hosted model, the docs point any OpenAI-compatible provider, Groq and OpenRouter among them, at the openai engine with a base_url. Engines whose API is not OpenAI-compatible, such as anthropic and vertexai, need a LangChain package and NEMOGUARDRAILS_LLM_FRAMEWORK=langchain.

yaml

models:
  - type: main
    engine: openai
    model: openai/gpt-oss-120b
    api_key_env_var: GROQ_API_KEY
    parameters:
      base_url: https://api.groq.com/openai/v1

Provider	Key	Variable	base_url and model
Groq	Free, console.groq.com/keys	`GROQ_API_KEY`	`https://api.groq.com/openai/v1`, `openai/gpt-oss-120b`
Gemini	Free, aistudio.google.com/apikey	`GEMINI_API_KEY`	`https://generativelanguage.googleapis.com/v1beta/openai/`, `gemini-2.5-flash`
OpenRouter	Free models end in `:free`, openrouter.ai/keys	`OPENROUTER_API_KEY`	`https://openrouter.ai/api/v1`, `google/gemma-4-31b-it:free`
OpenAI	Paid, platform.openai.com/api-keys	`OPENAI_API_KEY`	Leave out `base_url` and `api_key_env_var`

export GROQ_API_KEY=gsk_...

The first time a message is matched to an intent with the real embedding model, it downloads, about ninety megabytes.

Try it yourself

Change model in the snippet to gemini-2.5-flash and work out the other two lines that must change with it.

PreviousNeMo Guardrails overview

Little by little, you're building something great.