Extract Text from PDF

About Extract Text from PDF

Text inside a PDF is not stored as sentences. It is a list of instructions — draw this glyph of this font at this position — and very often the file has no notion of lines, paragraphs or reading order at all. This tool reads the text runs with pdf.js, together with their positions and sizes, and rebuilds lines, columns, paragraphs and headings from the geometry alone.

So two things decide the quality of what you get. First, whether the PDF records what each glyph means in Unicode. Second, whether the layout can be inferred from coordinates. Pages with no text layer at all, such as scans, come back empty, and the tool points you to OCR.

Why extracted text comes out as gibberish

Embedded fonts are usually subsets containing only the glyphs used, and the page content refers to glyph numbers inside that font rather than to characters. The table that maps those numbers back to Unicode is the ToUnicode CMap. When it is missing or wrong, a page can look perfect on screen while copy-paste yields random symbols, accented junk or blank boxes — the shapes can be drawn without knowing what they mean. Ligatures are a milder case: “fi” drawn as one glyph may come out as a single ligature character, or not at all.

Type 3 fonts define each glyph with ordinary PDF drawing operators, and they often carry no Unicode information whatsoever. They turn up in output from older TeX setups, some CAD and reporting tools, and odd converters. No extractor can recover text that the file never recorded; rendering the page and running OCR is the practical route. A quick test: copy a sentence in your PDF viewer and paste it into a plain-text editor. If it breaks there, extraction will break too.

Why the order gets scrambled, and how this tool orders text

A PDF may draw its text in any order. Some two-column papers draw the first line of the left column, then the first line of the right, and so on; others draw headers and footnotes last. Following the file's order interleaves the columns line by line, which is the classic symptom of naive extraction.

This tool ignores the file's order and recomputes it with a recursive XY-cut. Glyphs are stacked into horizontal slabs; when several slabs share an empty vertical gutter at the same position, that region is treated as multi-column, split at the gutter, and each column is processed the same way. Output runs region by region from the top, and left to right across columns within a region. A figure caption that spans both columns, or a table that crosses the gutter, breaks the gutter and can put parts out of order.

Rotated and vertical text is taken out of the column analysis and read separately. Sideways table headers and 90° drawing labels are grouped by angle and read as lines in their own direction; vertical Japanese is read column by column from the right, top to bottom. Both are added as separate paragraphs after the page's horizontal text, not put back into their original spot (such as a table cell).

How spaces and line breaks are restored

Many PDFs never store a space character; they simply move the pen. The tool inserts a space when the gap between two glyphs exceeds 0.12 times the font size. Between Korean, Chinese or Japanese characters, whose glyphs are wide by design, it waits for 0.22 times. Letter-spaced headings can therefore come out as “H E A D I N G”.

Glyphs whose baselines agree within 35% of the font size form one line. With “join line breaks” on, end-of-line hyphenation such as “exam-/ple” is undone only when the joined word (“example”) appears elsewhere in the document or the second part is a plain suffix such as -tion or -ment; otherwise the hyphen is kept, so a compound split at a line end (“well-/known”) stays “well-known”. Chinese or Japanese lines are joined without a space.

Headers, page numbers and headings

Lines in the top and bottom 12% of each page are compared with digits replaced by #, and any line that recurs on at least half of the pages (minimum two) is removed as a header or footer. Stand-alone page numbers such as “- 12 -”, “12 / 40” or “Page 12” go too. A heading is a block at least 1.15 times the document's body size, no longer than three lines or 160 characters; heading sizes are ranked to give Markdown levels 1 to 4.

Worked examples

How to use it

1. Add a PDF; the tool reports columns per page and pages without text.
2. Choose TXT or Markdown, line-break handling and header removal.
3. Check the order in the preview, then copy or download.

Frequently asked questions

Related guides

• Making Scanned PDFs Searchable: How OCR Works and Which Scan Settings Matter