NeMo Guardrails 0.8: NVIDIA's framework for adding safety rails to any LLM

Adding security guardrails to an LLM in production is traditionally ad-hoc work, costly and difficult to maintain. NVIDIA's NeMo Guardrails offers a standardized framework for doing this on any model, without modifying its weights.

Version 0.8 introduces Colang 2.0, a declarative language specifically designed to define the permitted and prohibited behaviors of an LLM in a readable and auditable way. Colang flows allow you to specify exactly how the system should react to certain types of input.

The framework works as an intermediate layer: it intercepts incoming requests, evaluates them against policies defined in Colang, and decides whether to let them through to the model, block them, or modify the response. This approach is model-agnostic: it works with GPT-4 via API, local Hugging Face models, and any LLM backend.

Native integration with LangChain and LlamaIndex makes it immediately usable in existing enterprise pipelines.