NeMo Guardrails

NVIDIA's guardrails library sits between application code and a language model, deciding what reaches the model, what comes back, and which actions it may trigger. Rules run at five points: on user input, in dialog flows, on retrieved RAG chunks, around tool calls, and on the final answer. Dialog behavior is written in Colang, a Python-flavored language for describing user intents and scripted bot replies; the older 1.0 syntax and the newer 2.0 one both load. A config folder holds a YAML file naming the models and active rails, the Colang files, and optional Python actions. Prebuilt rails cover jailbreak and injection detection, masking of personal data, self-checks for facts and hallucinations, NVIDIA safety models for content and topic control, and outside moderation services. Install it with pip on Python 3.10 to 3.13, then call it from Python, launch the bundled server with its OpenAI-style chat completions API, or build the Docker image. Code is Apache 2.0. Anonymous telemetry is on by default and an environment variable disables it. Aimed at teams shipping chatbots or RAG assistants under enforceable policies.