Hello. I have some PDF parsing pipelines that need as high as possible accuracy to output extracted data from the PDF files. I implemented structured outputs using the python genai library with a Pydantic schema and application/json output format. It works.
However, when asking Gemini to suggest improvements it said that using structured outputs like this wasn’t as reliable as sending the json schema in the prompt and then parsing/verifying the JSON myself. Is this true, or is Gemini being pessimistic about its abilities? 🙂
LinkedIn
Twitter
