Have you tried BAML (https://github.com/boundaryml/baml)? It's really good at structured output parsing. We integrated it directly into our pipeline builder.