pdf

Read the PDF directly and the scanned pages come through OCR, giving you per-page markdown next to page images. For long documents pass a page range like 1-5, since one read covers at most 20 pages.

Load the file with pdf-lib, call getForm(), and list all field names first with getFields() before setting values. After filling, call flatten() to bake values in, then render the result and visually check that no text overflowed or landed in the wrong field.

Use qpdf: run qpdf --empty --pages a.pdf b.pdf -- merged.pdf to merge, or qpdf input.pdf --pages . 1-5 -- out.pdf to split. The same operations are available in JavaScript through pdf-lib's copyPages API.

Use both, page by page. Text extraction is exact for strings but blind to struck-through clauses, stamps, and merged cells; the image shows layout but is unreliable for long numbers. When they disagree, trust the image for layout and the text for exact strings.

Author the content as a .docx file and convert it with LibreOffice using soffice --headless --convert-to pdf. This gives real typography, headers, footers, and page numbers; reserve pdf-lib for manipulating existing PDFs rather than layout from scratch.

Field names can be misleading and text often overflows its box while the code reports success. Always render the filled PDF with cdx-render and inspect every page visually before delivering it.