IMHO LLMs cannot provide statistically confident measures, and they are terrible at pretending to be capable of doing so. What worked: You use an OCR that prov

IMHO LLMs cannot provide statistically confident measures, and they are terrible at pretending to be capable of doing so. What worked: You use an OCR that provides character/word-level bounding boxes and let the LLM extract from data. Then the LLM is capable of "calculating" a confidence of extracted data.