Guardrails Input And Output

Config has before_request_hooks=[pii_redact, prompt_injection_detect]. Question: order of evaluation and short-circuit behavior.

Hooks execute in declared order; first hook with action=deny short-circuits the request and the gateway returns the guard's denial response. Verify by sending a payload that fails the second hook only — observe it reaches the second hook because the first passed.

Input guard pii_redact with action='sanitize'; user sends 'my email is alice@example.com — please draft a reply'.

Guard rewrites the user content with the email redacted (e.g., '<EMAIL>') before the model sees it. Verify in the trace log that upstream received the sanitized payload, not the raw one. Surface a warning to the operator that PII was sanitized — do not silently rewrite user-visible content.

PII guard redacts 'OpenAI API' as a credit-card-looking pattern; user-visible response now reads '<REDACTED> API'.

Treat guard false positives as a calibration problem: maintain a regression set of legitimate user inputs and measure FPR over each guard version. Tune regex / model-based detectors and consider allowlists for known-good tokens. Do not disable the guard wholesale.