NeMo Guardrails
Open-source toolkit for adding programmable safety rails to LLM apps.
About
NeMo Guardrails is NVIDIA's open-source toolkit for putting programmable guardrails between application code and the LLM. Developers define input rails (jailbreak and injection blocking), dialog rails (topic boundaries and conversation flow via the Colang modeling language), and output rails (response verification, fact-checking, grounding) that run at runtime, independent of the underlying model.
Because rails are user-defined and interpretable rather than baked into model weights, teams get auditable control — at the cost of a custom DSL with a learning curve and extra LLM calls per message that add latency. It integrates with LangChain and LlamaIndex.
What it can do
- Input, dialog, and output rails
- Colang conversational flow language
- Jailbreak and prompt-injection blocking
- Fact-checking and retrieval grounding
- LangChain and LlamaIndex integration
Setup
- Run
pip install nemoguardrails(Python 3.10+). - Create a config with
general.instructionsand Colang flows defining allowed behavior. - Add input/output rails for jailbreak detection and topic control.
- Wrap your LLM calls with the guardrails runtime.
- Tune rail strictness against your eval set.
Prerequisites
Python 3.10+, an LLM provider API key.
Use cases
Works with
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: NVIDIA open-source toolkit; Apache 2.0; peer-reviewed paper (arXiv 2310.10501); v0.24.1 released; integrates with LangChain, LlamaIndex, Zscaler AI Guard; Colang DSL has a learning curve; rail checks add per-message latency; Runtime enforcement rather than model alignment — auditable and interpretable. License: Apache 2.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · NVIDIA open-source toolkit; Apache 2.0; peer-reviewed paper (arXiv 2310.10501)
- · v0.24.1 released; integrates with LangChain, LlamaIndex, Zscaler AI Guard
- · Colang DSL has a learning curve; rail checks add per-message latency
- · Runtime enforcement rather than model alignment — auditable and interpretable
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.