[lxgr]

How do you think their implementation works under the hood? I'm almost certain it's also just a variant of "structured outputs", which many inference providers or LLM libraries have long supported.