For the complete documentation index, see llms.txt. Markdown versions of all docs pages are available by appending .md to any docs URL.
Multi-layered guardrails
Chain multiple prompt guards so each request passes every check in order, for defense in depth.
You can configure multiple prompt guards that run in sequence, creating defense-in-depth protection. Guards are evaluated in the order they appear in the configuration.
Shared guardrails under llm.policies.guardrails apply to every model. Model-specific guardrails under llm.models[].guardrails are merged in for the selected model, so the effective policy is the shared baseline plus any model-specific checks.
Example configuration that combines shared and model-specific guardrails:
# yaml-language-server: $schema=https://agentgateway.dev/schema/config
llm:
policies:
guardrails:
request:
# Shared layer 1: Fast regex check for known patterns
- regex:
action: reject
rules:
- builtin: ssn
- builtin: creditCard
- builtin: email
rejection:
body: "Request contains PII and cannot be processed"
response:
- regex:
action: mask
rules:
- builtin: ssn
- builtin: creditCard
models:
- name: "*"
provider: openAI
params:
model: gpt-3.5-turbo
apiKey: "$OPENAI_API_KEY"
guardrails:
request:
# Model layer: OpenAI moderation for harmful content
- openAIModeration:
model: omni-moderation-latest
policies:
backendAuth:
key: "$OPENAI_API_KEY"
rejection:
body: "Content blocked by moderation policy"
# Model layer: Custom webhook for domain-specific checks
- webhook:
target:
host: content-safety-webhook.example.com:8000
response:
- webhook:
target:
host: content-safety-webhook.example.com:8000In this example, the shared policy catches obvious PII across every model, while the selected model adds moderation and webhook checks on top. The response-side regex masking from the shared policy still applies to the model.
