NVIDIA NeMo Guardrails
What is NVIDIA NeMo Guardrails?
NVIDIA NeMo Guardrails is an open-source Python library for developers that adds programmable safety checks around LLM inputs, outputs, and retrieved content. It combines YAML and Colang with Content Safety, Jailbreak Protection, Topic Control, PII Detection, and Hallucinations & Fact-Checking, and supports Custom Actions plus third-party APIs. The docs point to GitHub, PyPI, and LangGraph integrations, and the package ships as versioned releases from 0.14.1 through 0.21.0.
Last verifiedHow we evaluate
At a glance
- NeMo Guardrails is best for developers who need to add policy checks around LLM inputs and outputs.
- Yes — The library includes a local API server with chat-completions endpoints such as /v1/chat/completions.
What does NVIDIA NeMo Guardrails do?
NeMo Guardrails handles LLM safety by inserting programmable checks around user inputs, model outputs, and retrieved content. It combines YAML and Colang configuration with guardrail types such as content safety, jailbreak protection, topic control, PII detection, and fact-checking so teams can shape what the model may say and when it should refuse. The library also supports custom actions and third-party APIs, letting developers wire policy logic into the application flow instead of bolting moderation on afterward. Behind the scenes, the open-source Python package is designed for development-time integration and portable configs that can move to the NeMo Guardrails microservice for Kubernetes deployment. NVIDIA positions it as part of the broader NeMo stack for building and scaling LLM applications, and the docs point to a local API server with chat-completions endpoints such as /v1/chat/completions. The project ships with versioned releases from 0.14.1 through 0.21.0, and the docs show integrations with GitHub, PyPI, LangGraph, and named customers or users of the ecosystem such as NVIDIA safety models and community guardrail providers.
Why use NVIDIA NeMo Guardrails?
- Portable YAML and Colang configs let teams develop locally and move the same guardrails into production deployment paths.
- The library covers multiple safety layers, so buyers can combine content moderation, jailbreak detection, topic control, and PII handling in one policy stack.
- Built-in support for third-party and NVIDIA guardrail providers gives teams flexibility to mix self-checking with external moderation services.
- A local API server with chat-completions endpoints makes it easier to slot guardrails into existing application patterns.
- The open-source Python package and PyPI distribution lower adoption friction for teams already shipping Python LLM apps.
Who is NVIDIA NeMo Guardrails for?
- LLM application developers who need programmable safety checks around prompts and responses.
- Platform engineers who want guardrails that fit into Python-based deployment workflows.
- AI product teams who need topic control and refusal behavior for customer-facing assistants.
- Security and compliance teams who need PII masking and jailbreak detection in model flows.
- Teams building agentic workflows who need controls around tool use and third-party APIs.
What are NVIDIA NeMo Guardrails's key features?
Content Safety
Filters unsafe prompts and outputs with integrations like Llama Guard, ActiveFence, and GCP Text Moderation to reduce policy violations before they reach users.
Jailbreak Protection
Detects jailbreak attempts using NemoGuard Jailbreak Detection NIM, Prompt Security, or Cisco AI Defense so applications can block prompt-injection attacks earlier.
Topic Control
Constrains conversations to approved subjects with NemoGuard Topic Control NIM and LangGraph, helping teams keep assistants on-policy and avoid off-field answers.
PII Detection
Finds and masks personal data with GLiNER PII, Microsoft Presidio, or Private AI, which helps teams protect sensitive text before logging or sending it onward.
Agentic Security
Adds guardrails around tool-using agents with integrations such as CrowdStrike AIDR and PolicyAI, reducing risky actions when an LLM calls external systems.
Hallucinations & Fact-Checking
Checks model claims against sources using AlignScore, Patronus Evaluate API, or Patronus Lynx, helping teams catch unsupported answers before users rely on them.
Custom Actions
Runs your own Python or API-based actions through the library's local API server and /v1/chat/completions endpoint, so guardrails can fit existing workflows.
Knowledge Base
Uses retrieved context and custom initialization to ground responses in your own data, which helps the model answer from approved sources instead of guessing.
What does NVIDIA NeMo Guardrails integrate with?
- GitHub
- PyPI
- ActiveFence
- AlignScore
- AutoAlign
- Cisco AI Defense
- Clavata
- Cleanlab
- CrowdStrike AIDR
- Fiddler
- GCP Text Moderation
- GLiNER PII
- GuardrailsAI
- Llama Guard
- Pangea AI Guard
- Patronus Evaluate API
- Patronus Lynx
- PolicyAI
- Presidio
- Private AI
- Prompt Security
- Trend Micro
- Llama 3.1 NemoGuard 8B Content Safety
- LlamaGuard
- Fiddler Guardrails
- NemoGuard Jailbreak Detection NIM
- NemoGuard Topic Control NIM
- NVIDIA GLiNER-PII
- Microsoft Presidio
- LangGraph
What are NVIDIA NeMo Guardrails's use cases?
LLM app safety checks
LLM application developers use NVIDIA NeMo Guardrails to wrap prompts and responses with programmable checks before output reaches users. They combine Content Safety with Jailbreak Protection to block risky requests, then use Topic Control to keep the assistant inside approved boundaries.
PII-safe customer assistants
Security and compliance teams use NVIDIA NeMo Guardrails to reduce exposure of sensitive data in customer-facing flows. They apply PII Detection to mask personal information and pair it with Hallucinations & Fact-Checking so support answers stay safer and more trustworthy.
Agent controls for tool use
Teams building agentic workflows use NVIDIA NeMo Guardrails to govern when an assistant can call tools or third-party APIs. With Agentic Security and Custom Actions, platform engineers can constrain risky actions while still letting the agent complete approved tasks.
Python deployment guardrails
Platform engineers use NVIDIA NeMo Guardrails inside Python-based deployment workflows to add safety without rewriting their stack. They integrate the library into their application and use Custom Initialization plus Knowledge Base to tailor behavior to their product and field.
How does NVIDIA NeMo Guardrails work?
- Integrate NeMo Guardrails Library into Your Application and connect the first chat flow or agent entry point. Use the local API server with /v1/chat/completions if you want a familiar interface for testing.
- Add Content Safety and Add Jailbreak Protection to define which prompts and responses should be blocked or redirected. Tune the rules so unsafe requests are refused before they reach the model.
- Control Topic Conversation and Detect and Mask PII to keep assistants on approved subjects while protecting sensitive data. Use these checks to shape customer-facing replies and reduce compliance risk.
- Add Agentic Security and Custom Actions for workflows that call tools or third-party APIs. Limit which actions the agent can trigger, then route approved steps through your own logic.
- Build Your Own or Use Third-party Guardrail Solutions, then connect Knowledge Base and Caching to refine behavior over time. Review outputs, update policies, and keep responses consistent as usage grows.
Frequently asked questions
What is NVIDIA NeMo Guardrails?
NVIDIA NeMo Guardrails is an open-source Python library for developers that adds programmable safety checks around LLM inputs, outputs, and retrieved content. It combines YAML and Colang with Content Safety, Jailbreak Protection, Topic Control, PII Detection, and Hallucinations & Fact-Checking, and supports Custom Actions plus third-party APIs. The docs point to GitHub, PyPI, and LangGraph integrations, and the package ships as versioned releases from 0.14.1 through 0.21.0.
What is NVIDIA NeMo Guardrails used for? Who is it for?
NVIDIA NeMo Guardrails is used for Content Safety, Jailbreak Protection, and Topic Control. It's built for LLM application developers, Platform engineers, and AI product teams.
Does NVIDIA NeMo Guardrails have an API and what does it integrate with?
The library includes a local API server with chat-completions endpoints such as /v1/chat/completions.
Explore other Agent Tools & Integrations
