What problem does it solve?
NeMo Guardrails solves the problem of unsafe or policy-violating behavior in LLM applications by enforcing runtime controls that validate inputs and outputs before harmful content reaches users.
Core Features & Use Cases
- Jailbreak and prompt-injection detection: Blocks attempts to bypass instructions (e.g., “ignore previous instructions / DAN” patterns) using programmable rails.
- Input/output validation and self-checks: Adds guard flows that evaluate toxicity, hallucination risk, and other quality signals before returning responses.
- Fact-checking and PII/tone safety controls: Supports retrieval-backed verification and sensitive-data filtering (e.g., with Presidio), improving trustworthiness in production.
- Use Case: Deploy a customer-support chatbot that refuses illegal requests, masks PII, detects jailbreak attempts, and verifies factual claims with retrieval-based checks.
Quick Start
Configure NeMo Guardrails with a refusal flow for illegal requests, then run the rails to generate a safe response instead of passing the unsafe prompt directly to your LLM.
