Guardrails
Checks around a model's input and output that constrain what it can do.
Overview
Guardrails are the validation applied before a request reaches the model and after it answers: blocking disallowed input, checking that output parses, refusing to act on a tool call that fails validation. It solves the problem that a model is probabilistic and the systems around it are not, so its output cannot be trusted as an interface boundary. It exists because the alternative is to trust the model, and the failure is silent when that trust is misplaced.
Technology Graph
How Guardrails connects to the rest of the stack.
Applied to
- AI Agent application AI Agents Guardrails used for AI Agent Medium ✓ source
- Red Teaming technique Safety & Alignment Guardrails used for Red Teaming Medium ✓ source