NVIDIA · 2024-06-15 · major
NeMo Guardrails — Programmable LLM Safety Rails
NVIDIA's toolkit for adding programmable safety rails to LLM apps. Define conversation flows, topic boundaries, and safety checks using Colang DSL.
Define what your LLM can and can't do using a simple dialogue DSL.
Key specs
| GitHub stars | 4k+ |
|---|---|
| Colang dsl | yes |
What is it?
NeMo Guardrails lets you define safety rails for LLM conversations using Colang, a simple DSL. You specify allowed topics, forbidden responses, and conversation flows — the system enforces them.
How does it work?
Write Colang files defining rails: 'if user asks about X, respond with Y' or 'never discuss Z'. The runtime intercepts LLM calls, checks against rails, and blocks/redirects as needed.
Why does it matter?
Instead of hoping your system prompt holds, you get deterministic guardrails. The LLM literally cannot respond outside defined boundaries.
Who is it for?
Teams building customer-facing LLM apps that need strict topic/safety control.
Try it
pip install nemoguardrails