Expose not only the generated inference type, but also the source JSON schema cc @coyotte508 - would it be feasible / desirable?