Turn any PDF or scan into structured data and Excel (Vision mode)

Most documents arrive as PDFs, scans or phone photos, and the data inside has to be typed into something else. This workflow reads a document that you have loaded into a Cadrix Document Store with the PDF loader in Vision mode, which renders every page and lets a vision model read it. That works for scans, screenshots, tables, handwriting-like fonts and Arabic as well as English. You type the list of fields you want, such as "invoice number, total, due date, customer name", and the workflow returns each one with its value, page, confidence and a quoted snippet from the document.

Accuracy comes from verification, not from trust. The model may only copy values it can quote. A code step then searches the document text for the quoted snippet and for the value itself. Anything that cannot be found, that has less than high confidence, or that conflicts with another part of the document is marked CHECK, and fields that are simply not in the document are marked NOT FOUND instead of being filled in. Requested tables are written one sheet each.

The result is an Excel file with a Fields sheet, any table sheets and a Review Notes sheet. Clean extractions finish straight away. If anything was flagged, a person opens the file, compares the CHECK rows with the original and approves or rejects it before the data is used. Vision reading uses your OpenRouter credits for every page, so keep one document per store and start with short documents. Because the reader is only as good as the stored text, re-upsert after replacing a document.