Which PDF extraction APIs preserve the original text?

For PDFs that already contain a text layer, read that layer instead of running OCR: PyMuPDF (AGPL or commercial licence) and Docling (MIT, runs locally, lossless JSON export) return the exact characters stored in the file. For scanned PDFs you need OCR, and the most faithful low-cost APIs are AWS Textract Detect Document Text ($1.50 per 1,000 pages for the first million) and Mistral OCR 4 ($4 per 1,000 pages, $2 batch, with bounding boxes and confidence scores). LLM-based parsers such as LlamaParse's agentic tiers are best for layout but can normalise or rewrite text, so verify them against the source.

Why do some PDF parsers change the text?

Generative parsers use a vision-language model to read the page and write Markdown. That model can fix typos, expand abbreviations, merge hyphenated words or drop repeated headers — useful for RAG, wrong for legal, financial or audit work where the exact wording matters. Deterministic text-layer extraction cannot invent characters.

How do I keep reading order and layout without changing the text?

Use a layout-aware tool that still copies characters from the text layer: Docling detects page layout, reading order and table structure and exports lossless JSON; PyMuPDF's get_text with sort=True orders text top-left to bottom-right, and its dict mode returns blocks, lines and spans with coordinates. For scans, Mistral OCR 4 returns typed blocks with bounding boxes in reading order.

What is the cheapest way to extract text from millions of PDF pages?

Text-layer extraction with PyMuPDF or Docling on your own machines costs only compute. For scanned pages, Textract Detect Document Text drops to $0.60 per 1,000 pages after the first million a month, and Mistral OCR 4 batch is $2 per 1,000 pages.

Quick Answer

Best PDF Extraction APIs That Preserve the Original Text

Published: October 9, 2026

The short answer

“Preserve the original text” has one reliable answer: don’t OCR a PDF that already has text. Most PDFs generated by software carry a text layer; reading it returns the exact characters, with no recognition errors and no model rewriting them. Use OCR only for scanned pages, and prefer OCR engines that report confidence so you can flag doubtful words. Prices are list USD, read October 9, 2026.

Tool	Method	Preserves exact text?	Layout / reading order	Price	Best for
PyMuPDF	Reads text layer (OCR via Tesseract optional)	Yes — exact characters	`sort=True`; blocks/lines/spans with coordinates	Free under AGPL; commercial licence from Artifex	Fast, high-volume digital PDFs
Docling	Text layer + layout models; OCR for scans	Yes on digital PDFs	Layout, reading order, tables, formulas; lossless JSON	Free, MIT; runs locally or air-gapped	RAG pipelines needing structure and fidelity
AWS Textract (Detect Document Text)	OCR	High fidelity; per-word confidence	Lines and words with geometry	$1.50 / 1,000 pages (first 1M), $0.60 after	Scanned pages at AWS scale
Mistral OCR 4	OCR model	High; confidence at page, block or word level	Bounding boxes, typed blocks, reading order; headers/footers separable	$4 / 1,000 pages, $2 batch	Multilingual scans (170 languages), citations
LlamaParse	Tiered: basic text to agentic LLM/VLM parsing	Basic tier yes; agentic tiers can normalise	Strong tables, charts, images; per-page JSON	1,000 credits = $1.25; basic from 1 credit/page; 10K free credits	Complex layouts where structure beats verbatim text

Picks by job

Digital PDFs (contracts, reports, exports): PyMuPDF or Docling. PyMuPDF is the fastest way to get the stored text out, page by page, and its dictionary mode gives every span with its font and position so you can rebuild layout yourself. Check the licence: AGPL unless you buy Artifex’s commercial licence. Docling, an LF AI & Data project under MIT, adds layout analysis, reading order and table structure on top of the text layer and exports Markdown, HTML or lossless JSON — a better default when the output feeds an LLM.

Scanned documents where wording matters: Textract or Mistral OCR 4. Both are recognition engines, not rewriters, and both expose confidence so a reviewer sees the uncertain words. Textract is the cheaper option at volume. Mistral OCR 4 (released June 2026) returns typed blocks — titles, tables, equations, signatures — with bounding boxes for source-grounded citations, and can return headers and footers separately instead of mixing them into the body.

Complex layouts for RAG: LlamaParse agentic tiers, accepting that the text is a model’s reading of the page. Its Auto Mode routes each page to the cheapest tier that handles it.

A fidelity check that takes five minutes

Pick 20 representative pages, including tables and footnotes.
Extract with your candidate and with PyMuPDF (text layer as ground truth for digital PDFs).
Diff the outputs after whitespace normalisation. Any changed word in a digital PDF is a fidelity bug.
For scans, hand-check the words the engine marked low-confidence.

The wider vendor landscape is in best OCR and document extraction APIs; for structured fields with page citations see best PDF OCR API for structured JSON.

Last verified: October 9, 2026.

The short answer

Picks by job

A fidelity check that takes five minutes

Sources