add-llm-guardrails

To add LLM guardrails, you wrap validation checks around model calls to intercept prompt injection, detect PII, and block toxic outputs before or after processing. Input guardrails screen requests before reaching the model, while output guardrails validate structured responses and prevent sensitive data egress.

The best way to validate structured LLM outputs is to apply output guardrails that check for toxicity, groundedness, and sensitive data egress after the model generates a response.

Yes, you can integrate guardrails-ai and nvidia-guardrails with existing agent applications by wrapping the guardrail libraries around your current model calls to enforce safety. This approach requires implementing validation and error handling capabilities to manage intercepted requests.

Yes, you need input and output validation for LLM apps in regulated enterprise environments to ensure regulatory compliance, prevent PII leakage, and block toxic or off-topic requests.

LLM input guardrails detect off-topic requests by validating the prompt against predefined safety criteria, stopping prompt injection, PII detection, and irrelevant queries before processing.

Limitations of using llm-guard include the need to implement custom validation and error handling capabilities to manage rejected outputs, ensuring groundedness and toxicity checks function correctly within your specific agent architecture.