A PDF table is now JSON. Are the amounts and footnotes still right?
A collected OpenRouter summary describes extracting tables with file input and a strict JSON schema. Before using the result, check row alignment, negative amounts, separators and footnotes against the PDF.
What to consider
First trace each value back to its row and column, then check whether normalization changed its meaning. The collected summary says OpenRouter’s tutorial sends a PDF and a strict JSON schema to a file-capable model, with field descriptions covering parenthesized losses, thousands separators and footnote markers. That offers an approach to try; schema compliance alone does not establish factual accuracy. Keep the original text alongside the normalized value until the interpretation is confirmed.
Preserve row and column meaning before normalizing numbers
Choose a table you can check manually and define the columns and their meanings. Names, periods, units and amounts should remain traceable to the original table. Plausible numbers alone are insufficient. Flag values whose source cell is unclear instead of asking the model to fill gaps.
Include repeated headers across pages, merged cells and blanks in your own sample, and check for column shifts. These are proposed checks, not tests the tutorial is known to have passed. The summary supplies no results for those layouts and cannot support a promise about arbitrary PDFs.
Make every numerical transformation explainable
Define separate rules for the parenthesized losses, thousands separators and footnotes mentioned in the summary. Keep the source text, the value intended for calculation and the origin of its unit. Distinguish blanks, zeroes and unreadable entries. You can then revise normalization without guessing what the earlier run changed.
If a footnote limits an amount’s scope or reporting period, carry that qualification with the data. Mark its location and have the user decide which notes must remain. A numeric target field is no reason to discard a meaningful qualification or invent a unit conversion.
Verify one table before connecting a bulk import
Check a small sample row by row and record omissions, column shifts, changed signs and notes requiring interpretation. Compare the output with a manually confirmed reference. Passing a format check establishes valid structure; mark the content ready for use only after checking its meaning.
This approach is a candidate for workflows with defined fields, retained source PDFs and someone responsible for checking. Keep the necessary human approval for data used in settlement or formal disclosure. This article uses the collected summary only; it did not run the tutorial, measure accuracy or verify file limits or API prices.
Source links
The links below were supplied with the collected records. This article offers independent analysis of their summaries; we have not individually checked the original sources or tested the products.