Inline guardrails

Block or flag model output through agentgateway before it reaches users.

Source: examples/guardrails. Read the guide for the concepts behind it.

support.yaml

# evalsi guardrails apply -f examples/guardrails/support.yaml --server ...
name: support-chat
project: support
description: PII stays out of prompts and answers; no credentials, no jailbreaks.
redact:
  - builtin: email
  - builtin: us-ssn
  - builtin: credit-card
  - pattern: 'ACCT-\d{6,}'
    replacement: '<ACCOUNT>'
evaluators:
  - secret-leak          # safety pack: blocks credentials in the content
  - ref: refusal         # recorded as a score; used in block_when below
block_when: >-
  scores["secret-leak"] < 1 ||
  (phase == "response" && labels[?"tier"].orValue("") == "free" && content.size() > 20000)
message: This request was blocked by the support-chat guardrail.
timeout: 1.5s
failure_mode: closed     # an evaluator error blocks; "open" lets content through
mode: enforce            # "audit" decides and reports, but blocks nothing