Document Tools

A converter has to infer paragraph boundaries from vertical gaps, reading order from position, and table structure from ruling lines or aligned whitespace. Single-column reports with ordinary body text convert cleanly. Two-column academic layouts, sidebars, footnotes and tables without visible borders are where the guessing shows: text lands in the wrong order, a table becomes a grid of text boxes, and hyphenated line breaks survive into the Word file.

Scanned PDFs are a different problem entirely, because they contain no text at all — only an image of a page. Running them through a normal converter produces an empty document. Send those through OCR first, then convert the result.

In almost every large PDF, embedded images account for the bulk of the file. Compression therefore means downsampling those images and re-encoding them, typically to 150 DPI for screen reading or 300 DPI for anything destined for print, plus a JPEG quality drop. Secondary savings come from subsetting embedded fonts to the glyphs actually used and dropping page thumbnails and unused objects.

This explains the case people find surprising: a 40-page text-only PDF barely shrinks. Its content streams are already Flate-compressed and there are no images to downsample, so there is nothing left to reclaim. If a file is mostly text and still enormous, the cause is usually a fully embedded CJK font or a stack of duplicated resources rather than the text itself.

OCR accuracy is decided before the file reaches the recogniser. Scan or export at 300 DPI — 200 is workable for large clean type, and anything under 150 costs you accuracy no engine can recover. Straighten the page: a two-degree skew measurably raises the error rate on small text. Grayscale usually beats colour for plain documents, and heavy JPEG artefacts around letter edges hurt more than a slightly larger file would.

Pick the language explicitly. An engine running in English mode will happily turn Turkish diacritics into the nearest ASCII letter, and mixed-language documents need every language selected up front. The output is a searchable PDF: the original page image stays visible while an invisible text layer sits behind it, so the document looks identical but can be searched, copied and indexed.

Drawing a filled box over a name hides it visually and changes nothing underneath. The text remains in the content stream, and anyone can recover it by selecting the page, running a text extractor, or opening the file in an editor. This has produced a long series of public embarrassments involving court filings and government releases.

Real redaction deletes the underlying content, then flattens the page so nothing can be reversed. Metadata deserves the same attention: author names, the originating software, revision timestamps and XMP records survive most workflows untouched, and a document assembled from several sources often carries all of their metadata. Redact, flatten, then check the document properties before sending.