PDF Text Extractor

In-Depth Guide

Extracting text from a PDF means walking the content stream — the sequence of text-showing operators (Tj, TJ, ', ") and text-state operators (Tf, Tm, Td, TD) defined in ISO 32000-2 section 9 — and reassembling positioned glyphs into a logical reading order. The trick is that PDFs were designed for presentation, not reading order: a two-column layout places glyphs in visual-flow order, which is not the reading order (left column all the way down, then right column). A good extractor clusters text fragments by baseline y-coordinate, sorts left-to-right inside each cluster, and heuristically rejoins hyphenated line breaks. For scanned PDFs (image-only, no text layer), the extractor has nothing to pull — OCR is required separately. FastTool's tool runs extraction locally via PDF.js so confidential contracts, medical reports, and legal filings produce usable plain text without a cloud round-trip.

Why This Matters

Lawyers search discovery bundles for keywords. Researchers count term frequencies in papers for literature reviews. Product managers paste PDF product requirements into a ticket system. Accessibility engineers feed text to screen-reader users. All of these need reliable plain text from the PDF, not a paragraph of mangled spacing and dropped ligatures. And all of them benefit from keeping the source PDF local — legal holds, HIPAA PHI, unreleased research, confidential internal specs — rather than feeding them to an external service whose log retention and training-data policy are never quite as tight as you would like.

Real-World Case Studies

• Discovery bundle search. A litigator extracts text from a 2,400-page discovery bundle to grep for 'gross negligence' and 'indemnification' across every document. The extractor produces 2,400 per-page text files in under two minutes. She pipes them through ripgrep, finds nine responsive documents, and cites them in her brief. Local extraction keeps the protective-order bundle from transiting any service that could be compelled or breached — a non-negotiable condition of the court's confidentiality order.
• Literature review corpus. A PhD researcher extracts text from 320 peer-reviewed papers for a systematic review. The extractor normalises ligatures (fi, fl, ffi), stitches hyphenated line breaks, and outputs clean UTF-8 text she can feed into her NLP pipeline for term-frequency analysis. The alternative — scraping text out of a web viewer or paying for a citation database — would have eaten the first two weeks of her project on mechanical data preparation.
• Accessibility remediation audit. A compliance engineer audits 400 customer-facing PDFs for WCAG 2.1 conformance. The extractor returns the logical text, which the engineer diffs against the PDF's alt-text and structural tags. Documents where the extracted text differs materially from the tagged order flag as remediation candidates. All processing is local, so customer PII embedded in those PDFs never transits the engineer's corporate gateway into an external service, satisfying GDPR Article 32.

Technical Deep Dive

The extractor parses each page's content stream, tracking the current text matrix, font, and font-size state via the stack of q/Q save/restore operators. For every text-showing operator, it records the glyph CIDs or codes, looks them up in the active font's ToUnicode CMap (ISO 32000-2 section 9.10) to recover Unicode code points, and computes each glyph's bounding box using the font's widths table. The resulting list of (codepoint, x, y, width) tuples sorts into reading order via a two-pass algorithm: first group by baseline y (fragments whose baselines differ by less than a tolerance fraction of the font size are the same line), then sort each line by x. Between consecutive fragments on the same line, a small positive gap becomes a space character and a large gap suggests a column break. Hyphenated line-endings — where a word breaks at the line wrap — rejoin if the next line's first token would form a dictionary word. Text not extractable from the ToUnicode map (a broken or missing CMap) falls back to encoding-aware recovery using the font's /Encoding entry. Output is UTF-8 with optional per-page or whole-document concatenation.