Skip to main content
SAICA-KG

Tools

NeMo Guardrails

6,028

Colang-based programmable guardrails for LLM applications.

DetectionPost-generationFully autonomousStableTier 2BPromptEnvironment
Maturity tier details

Computed tier: 2B(raw score 4.01). Effective tier: 2B.

Facets:documented_in=0evaluated_on=0integration_surfaces=0.5paper_citations=0recent_commit=2stars=1.51

Source of truth:pipeline/maturity/score.py

Description

NVIDIA's open-source toolkit for adding programmable rails around LLM applications. Uses Colang, a DSL for specifying conversation flows, topical rails, safety rails, fact-checking rails, and jailbreak detection. Rails execute around the LLM call as input checks, dialog flow control, and output checks. Integrates with LangChain and supports pluggable policy models.

Why it's included

Occupies detection × post_generation × fully_autonomous with a policy-language approach distinct from Pydantic-schema validators. Acts primarily as output/flow moderation, not constrained decoding — hence post_generation rather than in_generation.

Security notes

Open-source (Apache-2.0) from NVIDIA. Colang rails can call external moderation/classifier models — audit those endpoints for data-residency. Rails themselves are code; treat rail bundles pulled from third parties as supply-chain surface. NVIDIA vendor-maintained; tracks NeMo stack versions.

Addresses failure modes

  • Scope creep — Actions outside declared task boundary.
  • Security vulnerability — Insecure patterns, injection flaws, weak cryptography, secret leakage.
  • Context pollution — Trajectory drift, hallucination spirals, overlong or off-task context.
  • Fabrication — Generating reference to non-existent APIs, packages, or symbols.

Techniques

colang-policy-dsltopical-railsfact-checking-railjailbreak-detectionoutput-moderation

View raw YAML on GitHub →