Tools 9 min read 21 October 2026

PDF Text Extraction: Why It Fails, How It Works, and Tools That Do It Right

Q: What is the best Python library for PDF text extraction in 2024?

For most use cases, PyMuPDF (via the fitz module) provides the best combination of speed, accuracy, and positional control. It handles complex encodings well, provides word/block/page-level extraction, gives coordinate data for each text element, and is actively maintained. For cases requiring the most detailed font and encoding inspection, pdfminer.six gives lower-level access. For the simplest code with good performance on well-formed PDFs, pypdf is lightweight and easy to use. For heterogeneous document pipelines (mixed PDF/Word/Excel), Apache Tika provides a single API. For the highest accuracy on production documents including complex encodings and scanned PDFs, Adobe Extract API (paid) is the most reliable but adds cost and a cloud dependency.

Extracting text from a PDF seems simple but regularly produces garbled output, missing characters, incorrect reading order, or empty results for scanned documents. This guide explains the five technical reasons PDF text extraction fails, covers the extraction APIs in PyMuPDF, pdfminer, pypdf, and Tika, and provides a diagnostic workflow.

Extracting text from a PDF seems like it should be trivial — the text is right there on the page. In practice, it regularly fails in ways that range from minor (a few garbled characters) to complete (an empty result for every page). Understanding why text extraction fails is the key to choosing the right tool and diagnosing problems when they occur. The fundamental issue is that PDF is a visual format, not a semantic one. A PDF page stores drawing instructions — paint this glyph at this position — not text content in any natural language sense. Text extraction is a reconstruction process: tools infer what characters were drawn, in what order, and assemble them into something resembling the original text. This reconstruction can fail at several distinct technical levels. This guide covers all five failure modes, the extraction APIs and tools that handle each case, and a diagnostic workflow for identifying which failure mode you are encountering.

Why PDF text extraction fails: five failure modes

1. Missing or incomplete ToUnicode CMap — The most common cause of garbled extraction. When a font uses custom glyph names or ligatures without a ToUnicode CMap, text extraction tools cannot reliably map glyph codes to Unicode characters. Result: missing characters, wrong characters, or complete gibberish. Especially common in PDFs from: old LaTeX/pdflatex, professional design software (InDesign with custom typefaces), and some PostScript-to-PDF workflows. 2. Scanned document without OCR text layer — A scanned PDF is an image; there is no text in the PDF content stream at all. Text extraction returns empty or nothing. Resolution: run OCR before extraction. See PDF OCR Guide for OCR options. 3. Incorrect or absent reading order — PDF text objects are stored in drawing order (the order the drawing instructions paint them), which may differ significantly from reading order for multi-column layouts, tables, figures with captions, and complex typographic arrangements. Extraction in drawing order produces text that jumps between columns, mixes figure captions with body text, and interleaves headers/footers with article content. 4. Text as outlines (converted to paths) — Some PDFs have text converted to vector paths — there is no font, no character code, just bezier curves that look like letters. This happens in: older publishing workflows, PDFs exported from some design tools with "create outlines" option, and watermarks or decorative text. Result: zero text extracted regardless of tool. 5. Image-masked text (white-on-white or invisible text) — Some PDFs place text with a rendering mode that makes it invisible (rendering mode 3 — invisible text) while keeping it in the content stream. This is used in OCR'd PDFs (hidden text layer behind image) and some DRM schemes. The text IS extractable, but the extracted content may not match what's visually visible.

Extraction tools: capabilities and use cases

PyMuPDF (Python, cross-platform) — The highest-quality general-purpose PDF text extraction library. Uses MuPDF's extraction engine, which handles: ToUnicode with fallback glyph name lookup; multi-column layout detection with page.get_text("blocks") or page.get_text("words"); reading order approximation; and coordinate-level access to extracted text. Best for: most extraction tasks, programmatic text analysis, and cases requiring position-aware extraction.

import fitz  # PyMuPDF
doc = fitz.open("document.pdf")
for page in doc:
    text = page.get_text("text")   # simple linear extraction
    blocks = page.get_text("blocks")  # position-aware block extraction

pdfminer.six (Python) — Provides the most granular control over PDF parsing: access to individual text objects, fonts, encoding, and coordinates. Slower than PyMuPDF but useful for diagnosing encoding issues. The LAParams (Layout Analysis Parameters) control how characters are grouped into words and lines. pypdf (Python) — Fast, simple extraction for well-formed PDFs with good ToUnicode support. Less robust than PyMuPDF for complex layouts. Use when speed matters for high-volume processing of well-formed documents. Apache Tika (Java, REST API) — Server-based extraction that handles PDF, Word, Excel, and hundreds of other formats with the same API. Useful for heterogeneous document processing pipelines. Tika wraps PDFBox for PDF extraction. pdftotext (poppler CLI) — The fastest command-line extraction tool. pdftotext -layout document.pdf - preserves layout with whitespace. Good for simple scripts; limited handling of complex encodings. Adobe Acrobat Extract API — Cloud API that uses Adobe's rendering engine for extraction. Handles encoding issues better than most open-source tools. Appropriate for production document processing where quality matters more than cost.

Diagnostic workflow for extraction failures

When text extraction produces wrong or empty output, follow this diagnostic sequence: Step 1: Is the PDF scanned? — Check if pages have any text operators in the content stream: python3 -c "import fitz; doc=fitz.open('x.pdf'); print(doc[0].get_text()[:100])" Empty result → likely scanned. Also check visually: select text in a PDF viewer. If nothing can be selected, it's a pure image PDF. Step 2: Check font encoding — Open the PDF in Acrobat Pro and go to File > Properties > Fonts. Look for fonts without "Embedded" status, or with unusual naming like ABCDEF+FontName (subsetting is fine). In pdfminer, enable debug output to see which fonts are being used and whether ToUnicode is present:

from pdfminer.pdfinterp import PDFResourceManager
# check font encoding during extraction

Step 3: Check reading order — Extract with layout preservation and compare to visual reading order. If blocks are out of order, try PyMuPDF's sort=True parameter on get_text (sorts by vertical position, then horizontal). Step 4: Check for outline text — If a page visually contains text but extraction returns nothing, the text may be outlines. In Acrobat Pro, try Edit > Select All — if text-looking elements are selected as paths (blue outline selection), they are vector outlines, not text. Step 5: Check text rendering mode — In pdfminer, inspect the rendering mode of text objects. Mode 3 is invisible text. If your extraction includes these invisible text objects, they may or may not correspond to visible content. Solution matrix: - Garbled characters → use PyMuPDF; if still garbled, the PDF needs regeneration with correct ToUnicode - Empty result, scanned → run Tesseract or cloud OCR first - Wrong order → use block-based extraction with position sort - Outline text → OCR the page image; no text extraction possible from paths

Frequently asked questions

Why does text I can select in Acrobat come out garbled when I extract it programmatically?

Acrobat has a fallback mechanism for missing or incorrect ToUnicode CMaps: it consults its internal font metrics and uses heuristics to reconstruct Unicode characters from glyph names and outlines. Open-source tools typically have less sophisticated fallbacks. The result: text that Acrobat can correctly copy-paste produces garbled output in pypdf, PyMuPDF, or pdfminer. The root cause is a missing or incomplete ToUnicode CMap in the embedded font. The definitive fix is to regenerate the PDF from the source document with a tool that produces correct ToUnicode data. As a workaround, extracting via Acrobat's command-line tool or Export > More Formats > XML leverages Acrobat's better fallbacks. Adobe's Extract API (cloud) applies the same Acrobat engine programmatically.

What is the best Python library for PDF text extraction in 2024?

For most use cases, PyMuPDF (via the fitz module) provides the best combination of speed, accuracy, and positional control. It handles complex encodings well, provides word/block/page-level extraction, gives coordinate data for each text element, and is actively maintained. For cases requiring the most detailed font and encoding inspection, pdfminer.six gives lower-level access. For the simplest code with good performance on well-formed PDFs, pypdf is lightweight and easy to use. For heterogeneous document pipelines (mixed PDF/Word/Excel), Apache Tika provides a single API. For the highest accuracy on production documents including complex encodings and scanned PDFs, Adobe Extract API (paid) is the most reliable but adds cost and a cloud dependency.

Can I extract text from a PDF with copy-paste protection?

Copy-paste protection in PDF is implemented via permission flags in the /Encrypt dictionary — specifically, the content copying permission bit. Tools that enforce permissions will refuse extraction; tools that ignore permissions (or are given the owner password) will extract normally. The PDF specification allows owners to restrict copying, but the content is still in the file — it is just flagged as restricted. qpdf can remove permission restrictions if you have the owner password: qpdf --decrypt --password=ownerpass input.pdf unrestricted.pdf. Important: whether removing permission restrictions is legally permissible depends on your jurisdiction, the terms of use of the document, and your purpose. For PDFs you own, it is always permissible. For PDFs you have received with restrictions, consult the document's terms of use.

Try it yourself

Everything in this article can be done directly in the free PDFTheGuide editor — no account required.

Edit PDF Annotate PDF Open editor