Indeed! Accuracy is only a part of the problem. One way to crack this is to maintain the layout in the extraction. Layout preservation means more context and better LLM interpretation. A write-up is here if you are curious https://unstract.com/blog/extract-table-from-pdf/