NeMo Guardrails
NVIDIA's open-source toolkit for implementing guardrails on LLMs. Uses configuration-as-code for defining conversational rules, handling harmful content, and preventing undesired behaviors with rail specifications.
Overview
NeMo Guardrails is NVIDIA's open-source toolkit for adding programmable guardrails to LLM-based conversational applications. It uses a configuration-as-code approach with "rails" (specific output control mechanisms) for topics like avoiding politics, following dialog paths, using particular language styles, and extracting structured data. The library includes built-in guardrails for input/output moderation, fact-checking, hallucination detection, jailbreak prevention, and integrates with NVIDIA NIM, OpenAI, Azure, Anthropic, and LangChain providers. Version 0.20 (January 2026) added an OpenAI-compatible server, IORails engine, and integrations with Cisco AI Defense.
The Verdict
Who Should Use NeMo Guardrails?
Best For
- Teams needing conversational flow control (dialog paths)
- Python developers wanting config-as-code guardrails
- NVIDIA NIM and GPU infrastructure users
- Applications requiring topic steering and style control
- Self-hosted deployments with custom rail logic
Not Ideal For
- Teams wanting managed cloud service (self-hosted only)
- Simple input/output filtering (may be overkill)
- Non-Python environments
- Teams without NVIDIA/GPU expertise
What's Great
- Free and open-source (Apache 2.0 license)
- Powerful dialog flow control with Colang DSL
- Built-in hallucination detection and fact-checking
- NVIDIA NemoGuard safety models integration
- OpenAI-compatible server for easy deployment
- PII detection via Presidio, Private AI, GLiNER integrations
- Streaming support and OpenTelemetry tracing
- Active NVIDIA maintenance and community
Watch Out For
- Steeper learning curve (Colang DSL)
- Self-hosted only (no managed cloud option)
- Python 3.10+ required (no 3.9 support)
- Best experience requires NVIDIA infrastructure
- More complex than simple guardrail libraries
Pricing
View all features & details
Rail Types
- Input rails (prompt filtering)
- Output rails (response filtering)
- Dialog rails (conversation flow)
- Retrieval rails (RAG filtering)
- Execution rails (tool use control)
- Topic rails (subject steering)
Safety Features
- Jailbreak detection
- Content safety (NemoGuard)
- Topic safety controls
- Hallucination detection
- Fact-checking integration
- PII detection & redaction
LLM Providers
- NVIDIA NIM
- OpenAI
- Azure OpenAI
- Anthropic Claude
- HuggingFace models
- LangChain (optional)
Integrations
- Cisco AI Defense
- Microsoft Presidio (PII)
- Private AI
- Guardrails AI Hub
- OpenTelemetry
- CrowdStrike AIDR
How It Compares
| Feature | NeMo Guardrails | LLM Guard | Lakera Guard | Guardrails AI |
|---|---|---|---|---|
| License | Apache 2.0 | MIT | Proprietary | MIT |
| Dialog Flow Control | Yes (Colang) | No | No | No |
| Config-as-Code | Yes | Python only | API | Python/YAML |
| NVIDIA Integration | Native | No | No | No |
| Self-Hosted | Yes | Yes | Enterprise | Yes |
| Managed Cloud | No | No | Yes | No |
| Learning Curve | Higher | Lower | Lower | Medium |
| Best For | Conversation control | Security | API security | Output validation |