Guardrails

Scans input or output for sensitive data like names, addresses, or account numbers. It masks or redacts this data to protect user privacy.

Input: Before prompt processing to prevent sensitive data logging. Output: Before display to ensure the model doesn't leak PII, even from its training data.

Compares the AI's generated response against verified, external sources (grounding) to check factual accuracy. Prevents the model from generating fabricated or incorrect information.

In Retrieval-Augmented Generation (RAG) systems or any application generating high-stakes, factual content (e.g., legal, financial, or medical summaries).

Analyzes the user's prompt for malicious intent, obfuscation, or manipulation tactics designed to bypass safety filters. Blocks attempts to force the model to perform disallowed or harmful actions.

On user-facing applications where the model's core instructions must be protected from external attacks to maintain system integrity and security.

Uses classifiers to flag and filter text (input and output) that is toxic, hateful, explicit, or discriminatory. Ensures all parts of the conversation are safe and compliant with ethical guidelines.

Input: To stop harmful prompts before they reach the LLM. Output: To filter any inappropriate content the model may inadvertently generate, protecting brand reputation.

Employs a separate, often smaller, language model to evaluate the primary LLM's final output against detailed rules for accuracy, style, and safety. Serves as an automated, sophisticated quality assurance layer.

As the final check in complex AI agents or pipelines that require the response to strictly adhere to multiple, nuanced constraints (e.g., tone, format, and compliance).