Empty cells are another useful row-level control. Some extractors omit them entirely, shifting the remaining values left and producing fluent text with the wrong column assignments. A row with a blank rate and a populated duration can reveal that failure more clearly than a fully populated row.
I would preserve explicit empty-cell positions in the representation, then check the same row after a page break or repeated header. That extends your acceptance record with a structural failure that can survive text extraction, while keeping the expected result grounded in the original cell coordinates.
Ahmet Özel
AI Engineer. Computer Vision, RAG and LLM agents.
Empty cells are another useful row-level control. Some extractors omit them entirely, shifting the remaining values left and producing fluent text with the wrong column assignments. A row with a blank rate and a populated duration can reveal that failure more clearly than a fully populated row.
I would preserve explicit empty-cell positions in the representation, then check the same row after a page break or repeated header. That extends your acceptance record with a structural failure that can survive text extraction, while keeping the expected result grounded in the original cell coordinates.