Source: NVIDIA NeMo Guardrails documentation
NVIDIA's docs and README for the open-source NeMo Guardrails library: five rail types, config.yml and Colang, self-check rails, streaming and parallel options.
Contents
- What: official library documentation and GitHub README (overview, “How it works”, guardrail catalog, configuration reference, Colang 1.0 input-rails tutorial)
- Publisher: NVIDIA
- Date: “latest” docs read 2026-11-14; latest released version 0.24.1 per the README (no release date given)
- Original: NeMo Guardrails docs · configuration reference · input rails tutorial · GitHub
Summary
NeMo Guardrails is a Python library (and a Kubernetes microservice using the same config format) that sits between an application and its LLM and runs “rails” at five points of a request. Rails are configured in YAML (config.yml, prompts.yml) and Colang flows, plus Python custom actions. It ships self-check rails (the LLM judges its own input/output with a yes/no prompt) and integrations with dedicated safety models and third-party services.
Key claims
- Five rail types, in request order: input (allow, alter or reject the user message), retrieval (filter or transform RAG chunks), dialog (Colang flows over canonical-form messages that pick the next step), execution (validate tool/action inputs and outputs), output (allow, edit or block the response).
- The runtime is event-driven: a user message becomes an event; depending on config it may generate a canonical user intent, decide the next step, run an action, retrieve chunks, call the LLM and generate the bot message. Dialog prompt tasks are
generate_user_intent,generate_next_steps,generate_bot_message. - Colang 1.0 and 2.0 are both supported; 1.0 is the default.
- Config keys:
models(typesmain,embeddings,jailbreak_detectionare reserved;content_safetyis a convention referenced as$model=content_safety),instructions,rails.input.flows,rails.output.flows,prompts,knowledge_base,tracing. - Self-check input: a separate LLM call with the
self_check_inputprompt; answer “Yes” blocks. In the tutorial’sexplain()output a blocked request costs 1 LLM call, an allowed one 2 (check +generalanswer). Missing prompt raises an error at load; reasoning models need an explicitmax_tokens(default fallback 1024). - Output-rail streaming (
rails.output.streaming):enabled(default false),chunk_size200 tokens,context_size50 tokens carried between chunks,stream_firsttrue (stream before the rails run on the chunk). rails.input.parallel/rails.output.parallel(default false) run rails concurrently;rails.dialog.single_call.enabledmerges intent and response into one LLM call, with fallback to multiple calls.- Catalog categories: content safety (e.g. Llama 3.1 NemoGuard 8B Content Safety, Llama Guard), jailbreak protection (heuristics, self-check, NemoGuard Jailbreak Detect NIM), topic control, PII (Presidio, GLiNER-PII and others), agentic security, tool-call rails, hallucination and fact-checking, LLM self-check, third-party APIs (ActiveFence, Cisco AI Defense, Guardrails AI, and more).
- Server mode exposes an OpenAI-style
/v1/chat/completions; LangChain and LangGraph integrations exist.