What breaks in document AI

Tables and layout

15 of 15 quotes

• “Two-column pages are read straight across, so sentences from both columns interleave. Ruled tables flatten into a stream of numbers with no row/column structure.”

• “The hybrid chunker then even breaks that single table apart into meaningless values.”

• “Right now, it's failing to detect some of the tables entirely and splitting others into pieces.”

• “The current version 2.0.0 handles the tables worse than previous versions”

• “5-level merged cell tables, bilingual Hindi+English content in the same cell, strike-through formatting, cross-referenced amendments.”

• Hit this one too? We are building for it.

• “I tried paid online services like extend, reducto, landing, gemini, none are 100% accurate”

• “Multi-column layouts (IEEE / academic 2-column style) are getting linearized incorrectly into the markdown output”

• “Multi-column PDFs, tables, headers/footers, broken reading order, scanned pages, repeated boilerplate — that was doing way more damage than we thought.”

• “I used `pdfplumber` to extract the text from the PDFs (which is a nightmare on its own)”

• “I didn't like how tables were links to separate markdown files.”

• “The extracted table contains misplaced column names and corrupt data parsing pipeline.”

• “The entire highlighted region in the PDF (routing, flight info, handling info, and goods details) is not extracted”

• “for some reasons it merges the rows with STATE UT VT together, causing the output to be wrong. this is not a scanned document.”

• “My main gripe with openwebui, in addition to it being slow is the fact that it mangles documents in the OCR step. tables that could have been understood great by an multi modal llm, just gets mangled by the ocr and lost”

• “since November, using these same settings has resulted in incor”

• Hit this one too? We are building for it.