NeMo Guardrails

NeMo Guardrails by NVIDIA is an open-source toolkit designed to help developers make large language model (LLM) applications—like chatbots and AI-powered agents—safe, trustworthy, and compliant. With NeMo Guardrails, you can easily add customizable “guardrails” that control what AI agents can do or say, enforce security policies, and keep conversations on track, all with minimal setup. It works with popular AI frameworks and models, providing a reliable safety net so AI agents avoid unsafe, off-topic, or inappropriate outputs, whether in customer service, research, or enterprise settings.

Key Features

• Programmable guardrails to control content, filter sensitive or unsafe topics, and prevent harmful outputs.
• Protection from LLM vulnerabilities like jailbreaks, prompt injection, and privacy leaks with content moderation and PII detection.
• Plug-and-play integration with frameworks like LangChain and support for major LLMs (OpenAI, Llama, Falcon, and more).
• Flexible dialog control and topic enforcement using the unique Colang language, letting you steer conversations and define custom flows easily.

Use Cases

• Enforce content safety and compliance in customer service or chatbot solutions.
• Moderate and protect question-answering or Retrieval Augmented Generation (RAG) applications for enterprise knowledge bases.
• Add layered security and reliability when using LLM endpoints or chaining multiple AI agents together.

Technical Specifications

• Written in Python, supporting Python 3.9–3.12, and requires basic C++ tools for installation.
• Ships as an open-source toolkit (Apache 2.0 license) and also as ready-to-use microservices for enterprise deployment.
• Guardrails can be configured at multiple levels: input, output, dialog, retrieval, and execution; supports async and sync operations through a simple API.