LLM Guardrails

Different sense of "guardrail"

This note is about constraining LLM inputs and outputs. The experimentation/governance sense used in Search Governance and A-B Testing for Search — a metric threshold that halts a rollout — is unrelated.

Definition

A guardrail is, in James Briggs’s phrasing, “a semi or fully deterministic shield” placed around a generative model: a mechanism that validates or constrains what goes in and what comes out, without relying on the model’s own judgement.

Eugene Yan lists guardrails as one of seven patterns for LLM systems — “to ensure output quality” — covering structural correctness, factuality, and safety.

What They Are Applied To

  1. Safety and topic guidance — filtering malicious inputs, keeping conversation on-subject
  2. Deterministic dialogue — fixed response flows for predictable queries
  3. Retrieval control — deciding by semantic similarity whether retrieval is needed at all
  4. Agent tool use — governing when external tools or APIs may be invoked

Points 3 and 4 make guardrails a retrieval concern, not only a safety one: the decision “should this query hit the index?” is the same decision Query Routing makes.

NeMo Guardrails and Colang

NVIDIA’s NeMo Guardrails is the most documented implementation. It is programmed in Colang, a modelling language for dialogue with three building blocks — user message blocks, bot message blocks, and flow blocks. User utterances are encoded into semantic vector space (MiniLM by default) and matched against predefined canonical forms, so most routing happens by embedding similarity without an LLM call.

Supporting pieces: utterances (example phrases per canonical form), variables ($name), and actions (Python functions exposed via register_action()).

NeMo Guardrails vs Llama Guard

Wenqi Glantz compared the two (February 2024). Llama Guard is a fine-tuned Llama 2 derivative acting as an input-output classifier with six built-in unsafe categories, extensible with custom ones. NeMo Guardrails is the broader programmable framework.

Result
Input-moderation accuracy (18 adversarial prompts)89% for both
HardwareNeMo runs on free-tier T4; Llama Guard needs A100

Equal accuracy at very different cost is the practical finding.

Articles