Your LLM Can Return Perfect JSON and Still Be Wrong

Using large language models to extract structured data like JSON can lead to subtle errors even when the output format is perfect. When source data is incomplete, a rigid schema can force the model to invent values for required fields, a process known as hallucination. To mitigate this, schema fields should be made nullable to accurately reflect missing information from the source text. Additionally, including an "evidence" field that quotes the source text for each extracted value helps verify the data's origin, while post-processing validation with tools like Pydantic ensures the data is logically sound.