Use pydantic-yaml to save on token costs?

[lingster]

In order to save on costs, could we implement a version of this that makes llm respond in yaml? This may also have a benefit of faster response times as less tokens would need to be generated and returned.

Perhaps this could be added as a flag?
The trick would be to ensure we can correctly validate the yaml is in a format that matches the pydantic model.

[jxnl]

this would work if we used mode=MD_YAML, happy to take a pr on this, and add it into the benchmarks