NVIDIA-NeMo/Guardrails

The toolkit uses a declarative configuration approach to define safety policies and conversation flows, with support for input/output rails, dialog path steering, fact-checking, and jailbreak protection through semantic similarity detection. It provides both Python API and async-first server modes, integrating seamlessly with LangChain chains and multiple LLM providers (OpenAI, open-source models), while managing complex interactions through configurable guardrail policies that mediate between applications and language models.

5,772 stars. Actively maintained with 16 commits in the last 30 days.