← Back to Glossary
Guardrails
Input and output validation layers that enforce safety rules, topic boundaries, and formatting constraints.
Why it exists
Prevents prompt injection attacks, PII leaks, inappropriate language, and malformed tool parameters before execution.
How it works
Runs classifier models, regex filters, or Pydantic validators on inputs and outputs before passing them onward.
Related terms
Where you'll use this
roadmap