Reasoning behind some of the JSON Schema restrictions

[amiasato]

I'd like to have some more insight on some restrictions that the new constrained decoding API has for JSON Schemas. More specifically, why are the following disallowed:

• Generic objects (even plain str->str maps are not valid fields)
• Examples (comprehensible place to employ few-shot prompting)
• Defaults

None seem to be limitations with CFG design, so I guess they are motivated by issues with the model. It would be super nice to have them possible to enable (maybe some warnings, disabled by default, idk).

Thanks

[tamish-max]

My understanding is that these restrictions mostly come from constrained decoding, not from JSON Schema itself.

The decoder wants a fairly closed grammar it can enforce token-by-token. That is why some schema features are awkward:

• open-ended maps / arbitrary object keys blow up the grammar state space;
• examples and default are helpful authoring hints, but they are not validation rules the decoder can strictly enforce;
• tighter object/array schemas are much easier to compile into deterministic decoding rules.

Workarounds that usually help:

• represent dict[str, T] as list[{key, value}];
• apply defaults in application code after parsing;
• put examples in the prompt/tool description instead of the schema.