How to Extract Clean Tables From PDFs for LLMs

Most questions people ask an AI about a document are questions about a table: a revenue figure, a date in a row, a value in a column. Tables are also exactly where PDF text extraction fails hardest, which is why LLMs so often return the wrong number.

Why PDF tables break

A PDF does not store a table as a table. It stores cell text as glyphs placed at coordinates. When you extract the text, the grid that gave those numbers meaning is gone.

• Cells flatten into a single run of numbers with no rows or columns.
• Multi-line cells split across lines and merge with their neighbors.
• Header rows detach from the data beneath them.
• Merged cells and nested headers collapse into ambiguity.

Feed that to an LLM and ask which quarter had a 31 percent margin, and it genuinely cannot tell which number belongs to which row. So it guesses, and the guess is often wrong.

What a good table extraction produces

The goal is a Markdown table where columns are aligned and headers sit on top. Every value stays tied to its column header and its row, so a question about a number retrieves the number with its label intact. That one change fixes most wrong-number answers.

How to do it

1. Detect the table region instead of reading the page top to bottom.
2. Reconstruct rows and columns from cell positions, not from line breaks.
3. Keep headers attached to their columns.
4. Output a Markdown table, which LLMs parse reliably.

This is hard to do well across the messy variety of real PDFs, which is why most home-grown extractors get tables almost right and then fail on the tricky ones: multi-column pages, merged cells, and tables that span page breaks.

The shortcut

PackForAI reconstructs tables from PDFs (and Excel, Word and more) as clean Markdown tables, with headers and real dates intact, and shows the token count. Upload a PDF with a table at packforai.com and check whether the numbers land in the right columns.

If the model gives you the wrong number from a table, the table was probably broken before the model ever saw it.