I wonder whether that could be fixed by using an approach like this [1] [2], to guarantee the model output meets a schema.

[1] https://news.ycombinator.com/item?id=37125118 [2] https://blog.normalcomputing.ai/posts/2023-07-27-regex-guide...