NeMo Guardrails

What it is

A Python library that wraps a large language model with rules that control what users can ask and what answers leave the system. Rails can block, rewrite, or route a message before the model runs, while it runs, or after it replies.

What it helps with

- Blocking jailbreaks and prompt injection attempts on the way in
- Checking answers for facts and hallucinations before they reach the user
- Steering a bot along a fixed dialog path with Colang definitions
- Filtering chunks pulled from a knowledge base in retrieval setups

Who it's for

AI engineers and security teams adding safety controls to LLM applications they build and run.

Worth knowing

A configuration can mix input, dialog, retrieval, and output rails, or run with no rails and simply forward requests to the model. Python 3.10 through 3.13 is required.

Does it work only with NVIDIA models?
No, it supports a range of LLM providers through its configuration.

Is there a research paper behind it?
Yes, an arXiv paper describes the system and its early evaluation.