See how the AI accuracy tests perform with a prompt that contains just the sample documents, no schema. 

This will not perform well with very polymorphic data, however, we currently don't parse that many documents in our schema generation anyway so that may not be a large factor.