PDF Oxide for Python — The Fastest PDF Toolkit for Python
The fastest Python PDF library for text extraction, image extraction, and markdown conversion. Powered by a pure-Rust core, exposed to Python through PyO3. 0.8ms mean per document, 5× faster than PyMuPDF, 15× faster than pypdf. 100% pass rate on 3,830 real-world PDFs. MIT / Apache-2.0 licensed.
Part of the PDF Oxide toolkit. Same Rust core, same speed, same 100% pass rate as the Rust, Go, JavaScript / TypeScript, C# / .NET, and WASM bindings.
Quick Start
pip install pdf_oxide
from pdf_oxide import PdfDocument
with PdfDocument("paper.pdf") as doc:
print(len(doc)) # number of pages
for page in doc:
text = page.text # lazy property
md = page.markdown(detect_headings=True)
Why pdf_oxide?
- Fast — 0.8ms mean per document, 5× faster than PyMuPDF, 15× faster than pypdf, 29× faster than pdfplumber
- Reliable — 100% pass rate on 3,830 test PDFs, zero panics, zero timeouts, no segfaults
- Complete — Text extraction, image extraction, search, form fields, PDF creation, and editing in one package
- Permissive license — MIT / Apache-2.0, unlike PyMuPDF (AGPL-3.0) — use freely in commercial and closed-source projects
- Pure Rust core — Memory-safe, panic-free, no C dependencies
- Native wheels — No build step, no system dependencies, no Rust toolchain required
Performance
Benchmarked on 3,830 PDFs from three independent public test suites (veraPDF, Mozilla pdf.js, DARPA SafeDocs). Text extraction libraries only. Single-thread, 60s timeout, no warm-up.
| Library | Mean | p99 | Pass Rate | License |
|---|---|---|---|---|
| PDF Oxide | 0.8ms | 9ms | 100% | MIT / Apache-2.0 |
| PyMuPDF | 4.6ms | 28ms | 99.3% | AGPL-3.0 |
| pypdfium2 | 4.1ms | 42ms | 99.2% | Apache-2.0 |
| pymupdf4llm | 55.5ms | 280ms | 99.1% | AGPL-3.0 |
| pdftext | 7.3ms | 82ms | 99.0% | GPL-3.0 |
| pdfminer | 16.8ms | 124ms | 98.8% | MIT |
| pdfplumber | 23.2ms | 189ms | 98.8% | MIT |
| markitdown | 108.8ms | 378ms | 98.6% | MIT |
| pypdf | 12.1ms | 97ms | 98.4% | BSD-3 |
99.5% text parity vs PyMuPDF and pypdfium2 across the full corpus. PDF Oxide extracts text from 7–10× more "hard" files than it misses vs any competitor.
Installation
pip install pdf_oxide
Pre-built wheels for Linux (x86_64, aarch64, musl), macOS (x86_64, arm64), and Windows (x86_64). Python 3.8 through 3.14. No system dependencies, no Rust toolchain required.
API Tour
Open a document
from pdf_oxide import PdfDocument
# Path can be str or pathlib.Path
doc = PdfDocument("report.pdf")
print(f"Pages: {len(doc)}")
print(f"PDF version: {doc.version()}")
# Context manager — closes automatically
with PdfDocument("report.pdf") as doc:
for page in doc:
print(page.text)
# From bytes or with a password
doc = PdfDocument.from_bytes(pdf_bytes)
doc = PdfDocument("encrypted.pdf", password="secret")
Page objects
PdfDocument is a sequence of Page objects. Pages are cheap to create; all
extraction is lazy and computed only when the property or method is accessed.
with PdfDocument("report.pdf") as doc:
print(len(doc)) # page count
# Iterate all pages
for page in doc:
print(f"Page {page.index}: {page.width:.0f}×{page.height:.0f} pts")
text = page.text # str
chars = page.chars # list[TextChar]
words = page.words # list[PyWord]
lines = page.lines # list[TextLine]
spans = page.spans # list[TextSpan]
tables = page.tables # list[Table]
images = page.images # list[dict]
annots = page.annotations # list[dict]
paths = page.paths # list[dict]
md = page.markdown(detect_headings=True)
html = page.html()
txt = page.plain_text()
# Render to PNG/JPEG bytes
png_bytes = page.render(dpi=150, format="png")
# Search within this page
hits = page.search("revenue", case_insensitive=True)
# Index access (negative indices supported)
first = doc[0]
last = doc[-1]
Text extraction (document-level)
text = doc.extract_text(0) # single page
all_text = doc.extract_text_all() # all pages joined
# Character-level
chars = doc.extract_chars(0)
for ch in chars:
print(f"{ch.char} at ({ch.x:.1f}, {ch.y:.1f}) size={ch.font_size:.1f}")
# Word-level
words = doc.extract_words(0)
for w in words:
print(f"{w.text} at {w.bbox}")
# Line-level
lines = doc.extract_text_lines(0)
for line in lines:
print(f"Line: {line.text}")
# Override the adaptive word/line gap thresholds (in PDF points)
words = doc.extract_words(0, word_gap_threshold=2.5)
lines = doc.extract_text_lines(0, word_gap_threshold=2.5, line_gap_threshold=4.0)
Format conversion (document-level)
# Markdown with optional heading detection and form-field inclusion
md = doc.to_markdown(0, detect_headings=True)
md_all = doc.to_markdown_all()
# HTML with optional CSS layout preservation
html = doc.to_html(0, preserve_layout=False)
html_all = doc.to_html_all()
# Plain text with automatic reading order
text = doc.to_plain_text(0)
text_all = doc.to_plain_text_all()
Scoped extraction
Extract content from a region of a page using within(). The region is (x, y, width, height) in PDF points.
header = doc.within(0, (0, 700, 612, 92)).extract_text()
region = doc.within(0, (50, 400, 500, 200))
region_words = region.extract_words()
region_images = region.extract_images()
Tables
tables = doc.extract_tables(0)
for table in tables:
print(f"Table with {table.row_count} rows")
Search
results = doc.search("quarterly revenue", case_insensitive=True)
for r in results:
print(f"Page {r['page']}: '{r['text']}' at {r['bbox']}")
# Single-page literal search
results = doc.search_page(0, "total", case_insensitive=True, literal=True)
Extraction profiles
Pre-tuned profiles adjust how raw text is parsed into words and lines for different document types.
from pdf_oxide import ExtractionProfile
words = doc.extract_words(0, profile=ExtractionProfile.form())
lines = doc.extract_text_lines(0, profile=ExtractionProfile.academic())
# Combine a profile with manual overrides
words = doc.extract_words(0, word_gap_threshold=1.5, profile=ExtractionProfile.aggressive())
Form fields
# Read all form fields
fields = doc.get_form_fields()
for f in fields:
print(f"{f.name} ({f.field_type}) = {f.value}")
# Fill and save
doc.set_form_field_value("employee_name", "Jane Doe")
doc.set_form_field_value("wages", "85000.00")
doc.set_form_field_value("retirement_plan", True)
doc.save("filled.pdf")
# Export form data as FDF or XFDF
doc.export_form_data("data.fdf")
doc.export_form_data("data.xfdf", format="xfdf")
Images
images = doc.extract_images(0)
for i, img in enumerate(images):
print(f"{img['width']}x{img['height']} {img['color_space']}")
img.save(f"image_{i}.png")
PDF creation
from pdf_oxide import Pdf, PdfBuilder, PageSize
# From Markdown, HTML, plain text, or images
Pdf.from_markdown("# Report\n\nHello **world**.").save("report.pdf")
Pdf.from_html("<h1>Invoice</h1><p>Total: $42</p>").save("invoice.pdf")
Pdf.from_text("Simple document content.").save("notes.pdf")
Pdf.from_image("photo.jpg").save("photo.pdf")
Pdf.from_images(["page1.jpg", "page2.png"]).save("album.pdf")
# Builder pattern for advanced control
pdf = (PdfBuilder()
.title("Annual Report 2025")
.author("Company Inc.")
.page_size(PageSize.A4)
.margins(72.0, 72.0, 72.0, 72.0)
.from_markdown("# Annual Report\n\n..."))
pdf.save("annual-report.pdf")
# Encryption
pdf = Pdf.from_markdown("# Confidential")
pdf.save_encrypted("secure.pdf", "user-password", "owner-password")
Async support
AsyncPdfDocument and AsyncPdf run all operations in a background thread, keeping your event loop free. Every method from the sync classes is available as an async counterpart.
import asyncio
from pdf_oxide import AsyncPdfDocument, AsyncPdf
async def main():
doc = await AsyncPdfDocument.open("report.pdf")
text = await doc.extract_text(0)
md = await doc.to_markdown(0, detect_headings=True)
pdf = await AsyncPdf.from_markdown("# Hello")
await pdf.save("hello.pdf")
asyncio.run(main())
OCR & Auto Mode
The published Python wheel ships with ocr built in. Install ONNX
Runtime, drop the models in PDF_OXIDE_MODEL_DIR, then let
pdf_oxide route per page (native text where present, OCR where the
page is image-only, graceful fallback when OCR is unavailable):
from pdf_oxide import PdfDocument
doc = PdfDocument("scanned-or-mixed.pdf")
text = doc.extract_text_auto(0) # recommended
For manual OcrEngine(det, rec, dict) usage,
doc.extract_text_ocr(page, engine), page-type classification, model
selection, and ONNX Runtime install recipes:
OCR Guide.
Other languages
PDF Oxide ships the same Rust core through six bindings:
- Rust —
cargo add pdf_oxide— see docs.rs/pdf_oxide - Go —
go get github.com/yfedoseev/pdf_oxide/go— see go/README.md - JavaScript / TypeScript (Node.js) —
npm install pdf-oxide— see js/README.md - C# / .NET —
dotnet add package PdfOxide— see csharp/README.md - WASM (browsers, Deno, Bun, edge runtimes) —
npm install pdf-oxide-wasm— see wasm-pkg/README.md
A bug fix in the Rust core lands in every binding on the next release.
Documentation
- Full Documentation — Complete documentation site
- Python Getting Started — Step-by-step Python guide
- Main Repository — Rust core, CLI, MCP server, all bindings
- Performance Benchmarks — Full benchmark methodology and results
- GitHub Issues — Bug reports and feature requests
Use Cases
- RAG / LLM pipelines — Convert PDFs to clean Markdown for retrieval-augmented generation with LangChain, LlamaIndex, or any framework
- Document processing at scale — Extract text, images, and metadata from thousands of PDFs in seconds
- Data extraction — Pull structured data from forms, tables, and layouts
- Academic research — Parse papers, extract citations, and process large corpora
- PDF generation — Create invoices, reports, certificates, and templated documents programmatically
- PyMuPDF alternative — MIT licensed, 5× faster, no AGPL restrictions
Why I built this
I needed PyMuPDF's speed without its AGPL license, and I needed it in more than one language. Nothing existed that ticked all three boxes — fast, MIT, multi-language — so I wrote it. The Rust core is what does the real work; the bindings for Python, Go, JS/TS, C#, and WASM are thin shells around the same code, so a bug fix in one lands in all of them. It now passes 100% of the veraPDF + Mozilla pdf.js + DARPA SafeDocs test corpora (3,830 PDFs) on every platform I've tested.
If it's useful to you, a star on GitHub genuinely helps. If something's broken or missing, open an issue — I read all of them.
— Yury
License
Dual-licensed under MIT or Apache-2.0 at your option. Unlike AGPL-licensed alternatives, pdf_oxide can be used freely in any project — commercial or open-source — with no copyleft restrictions.
Citation
@software{pdf_oxide,
title = {PDF Oxide: Fast PDF Toolkit for Rust, Python, Go, JavaScript, and C#},
author = {Yury Fedoseev},
year = {2025},
url = {https://github.com/yfedoseev/pdf_oxide}
}
Python + Rust core | MIT / Apache-2.0 | 100% pass rate on 3,830 PDFs | 0.8ms mean | 5× faster than the industry leaders
Metadata
Release files for pdf-oxide 0.3.78
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| pdf_oxide-0.3.78.tar.gz | 6.9 MB | Without provenance | Details for pdf_oxide-0.3.78.tar.gz |
Built distributions (wheels)
Total release size: 95.9 MB
Release files / pdf_oxide-0.3.78.tar.gz
| Download URL | pdf_oxide-0.3.78.tar.gz |
|---|---|
| Size | 6.9 MB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
18fb43c7b4e390400aab51f546a7b8c6ee471b5cbec88f4cf3dd74f6c43c49f2
|
|
BLAKE2b-256 checksum How to use checksums |
9952202c81a0ff60e64309dd7fc4682029f88d3fb07c827f4eaf78b53a7552ee
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Release files / pdf_oxide-0.3.78-cp38-abi3-win_arm64.whl
| Download URL | pdf_oxide-0.3.78-cp38-abi3-win_arm64.whl |
|---|---|
| Size | 10.8 MB |
| Tags | CPython 3.8 Windows ARM64 abi3 |
|
SHA-256 checksum How to use checksums |
f05e894c5d984284f56694f03a9351ea18d8c1ea081b4445007dfaa5c11dce85
|
|
BLAKE2b-256 checksum How to use checksums |
06cbebd9574ec3a984b347347dc0e9c370ae38ad26dfafcb99e2253c9823bb8b
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Release files / pdf_oxide-0.3.78-cp38-abi3-win_amd64.whl
| Download URL | pdf_oxide-0.3.78-cp38-abi3-win_amd64.whl |
|---|---|
| Size | 11.6 MB |
| Tags | CPython 3.8 Windows x86-64 abi3 |
|
SHA-256 checksum How to use checksums |
3c2800d15e60ce1d2ec5de5ff9a605c167e0c9964bb9f6143862d282ef6b46d0
|
|
BLAKE2b-256 checksum How to use checksums |
9561700cb40f9806df302e085a2d3b1133c21a95a54f2375bad3a04c419223ca
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Release files / pdf_oxide-0.3.78-cp38-abi3-musllinux_1_2_x86_64.whl
| Download URL | pdf_oxide-0.3.78-cp38-abi3-musllinux_1_2_x86_64.whl |
|---|---|
| Size | 11.7 MB |
| Tags | CPython 3.8 Linux musl 1.2+ x86-64 abi3 |
|
SHA-256 checksum How to use checksums |
e290619267875256d310ea692f2a748ddb9575d3c35ade61db51f3aaade1ddac
|
|
BLAKE2b-256 checksum How to use checksums |
5914b0a37b598af5f57bea25e130e1f3c1bbac549ec6dd00960c14d951e865ea
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Release files / pdf_oxide-0.3.78-cp38-abi3-musllinux_1_2_aarch64.whl
| Download URL | pdf_oxide-0.3.78-cp38-abi3-musllinux_1_2_aarch64.whl |
|---|---|
| Size | 11.0 MB |
| Tags | CPython 3.8 Linux musl 1.2+ ARM64 abi3 |
|
SHA-256 checksum How to use checksums |
119c45cc16e1763440e438ab3e2d0a59ff406ed4c5bf6b78716bd2dfb2083b05
|
|
BLAKE2b-256 checksum How to use checksums |
9168f3b875b44c844970ca7bb727ea417d313a89c176b0d43733267102b871f7
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Release files / pdf_oxide-0.3.78-cp38-abi3-manylinux_2_28_x86_64.whl
| Download URL | pdf_oxide-0.3.78-cp38-abi3-manylinux_2_28_x86_64.whl |
|---|---|
| Size | 11.5 MB |
| Tags | CPython 3.8 Linux glibc 2.28+ x86-64 abi3 |
|
SHA-256 checksum How to use checksums |
0b7e03008ca441e9183ef8f3c5d895424aca8119d5a860a205e1d16649f4f6a8
|
|
BLAKE2b-256 checksum How to use checksums |
df6c6b5957c884fb73582851810b5f1e4660da1d414c692ab77c2e784aecd9e0
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Release files / pdf_oxide-0.3.78-cp38-abi3-manylinux_2_28_aarch64.whl
| Download URL | pdf_oxide-0.3.78-cp38-abi3-manylinux_2_28_aarch64.whl |
|---|---|
| Size | 10.8 MB |
| Tags | CPython 3.8 Linux glibc 2.28+ ARM64 abi3 |
|
SHA-256 checksum How to use checksums |
dda4a7fb93f3b68a4b9bf52d0d314891d777c417f41993a0c2d8937325a59f2b
|
|
BLAKE2b-256 checksum How to use checksums |
1ce94774d28d8a0325809e3c4b1903d777c761b9058fa5dedd47c08caf1f5253
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Release files / pdf_oxide-0.3.78-cp38-abi3-macosx_11_0_arm64.whl
| Download URL | pdf_oxide-0.3.78-cp38-abi3-macosx_11_0_arm64.whl |
|---|---|
| Size | 10.5 MB |
| Tags | CPython 3.8 abi3 macOS 11.0+ ARM64 |
|
SHA-256 checksum How to use checksums |
538a809d35d9dec07e2e7a7568352bcf4114a8fb66d12f7452c802b2304d6946
|
|
BLAKE2b-256 checksum How to use checksums |
9a51cf844ffab595c22d181f219e786d7c0f80e8d2bfbee8d15a0f1b0cc88406
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Release files / pdf_oxide-0.3.78-cp38-abi3-macosx_10_12_x86_64.whl
| Download URL | pdf_oxide-0.3.78-cp38-abi3-macosx_10_12_x86_64.whl |
|---|---|
| Size | 11.1 MB |
| Tags | CPython 3.8 abi3 macOS 10.12+ x86-64 |
|
SHA-256 checksum How to use checksums |
a0c98d7753ec8593bd7e79ee41ad813b22a55fdd321cae85138e791802eac92b
|
|
BLAKE2b-256 checksum How to use checksums |
7ae0955517da3f8249239573ecafd756f7c647984b6c2a7c67701333556cf3ce
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Security for pdf-oxide 0.3.78
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.