Azure Content Safety, Prompt Shields and Language can act as the guardrails on a Gravitee LLM proxy through HTTP callouts, with no guardrail model loaded in the gateway
Pre-release buildInternal only
Made for Posten Bring
Say while it plays
The judging happens in Azure, so no model loads into the gateway container: gateway memory stayed at about 1.5 GB with the chain on or off.
One flow on an existing LLM proxy: a Groovy step shapes every turn, three HTTP callouts go to Language and Content Safety, a second Groovy step masks or blocks, and a native Interrupt policy returns the 403 so the model is never called.
The application needs no change: a LangChain agent only had its base URL pointed at the route, and policy lives in the gateway, not in each agent.
Don't claim
Don't claim reliable Finnish coverage: the recording says it is weaker and Microsoft's documentation backs that up.
Don't claim harm is re-scored on earlier turns: the recording says it sometimes isn't re-scored the way it is on the first turn.
Don't present this as native Gravitee functionality: it is a workaround of callouts and Groovy that a native policy would replace.
Chapters
Live demo: multilingual replies and content blocking
Azure services and policy configuration in Gravitee Gateway