pylopdf

PDF editing, rendering, extraction, and generation for Python — a pymupdf-style API on a pure-Rust core, MIT licensed, with no mandatory Python dependencies.

Documentation · Getting started · API reference · pymupdf migration guide · Benchmarks

pylopdf combines the 2026 Rust PDF ecosystem behind one Python API: lopdf for editing, hayro (the pure-Rust PDF renderer adopted by Typst) for rendering and positioned extraction, and krilla with HarfRust for shaped, subset-embedded text and form appearances.

Why pylopdf?

pylopdf's range covers the five native abi3 wheels built from the 0.14.0 sources with the thin-LTO release profile on GitHub-hosted runners on 2026-10-17; the published 0.14.0 wheels are 6.8–7.9 MiB. The other ranges are published files for pymupdf 1.28.0, pypdfium2 5.12.1, pdf-oxide 0.3.75, and pikepdf 10.10.0 on 2026-07-25.

• AGPL-free: covers the common pymupdf use cases under the MIT license
• Small and self-contained: fits size-constrained environments such as AWS Lambda and Cloudflare Workers
• One wheel per platform: abi3 covers CPython 3.11–3.15, plus native cp314t and cp315t wheels for free-threaded Python
• Familiar: the API is modeled after pymupdf, with a migration guide

Install

pip install pylopdf

Optional extras:

pip install "pylopdf[jp]"    # Japanese rendering fallback and text generation
pip install "pylopdf[ko]"    # Korean
pip install "pylopdf[zh-cn]" # Simplified Chinese
pip install "pylopdf[zh-tw]" # Traditional Chinese
pip install "pylopdf[ar]"    # Arabic generation (also: he, hi, th)
pip install "pylopdf[ocr]"   # offline PP-OCRv6 recognition, pure Rust,
                             # no system executables or network at runtime

The locale-specific CJK extras provide both non-embedded PDF fallback and generation fonts. The ar, he, hi, and th extras provide generation fonts for Arabic, Hebrew, Devanagari, and Thai. Choose jp, ko, zh-cn, or zh-tw explicitly for the required regional glyph conventions; there is no generic CJK font extra.

Building from source requires a Rust toolchain: uv sync.

Quickstart

import pylopdf

doc = pylopdf.open("input.pdf")              # or pylopdf.open(stream=pdf_bytes)

# Extract: plain text, positioned words, search, tables, Markdown
text = doc.get_page_text(0)
words = doc[0].get_text("words")             # (x0, y0, x1, y1, word, block, line, word_no)
hits = doc[0].search_for("tax")              # case-insensitive, list[Rect]
tables = doc[0].find_tables()                # bordered grids incl. merged cells
markdown = doc.to_markdown()                 # RAG / LLM preprocessing

# Render: PNG / SVG / raw pixels
png: bytes = doc.render_page(0, dpi=300)
batch = doc.render_pages([0, 1, 2], scale=2, workers=4)
svg: str = doc.render_page_svg(0)
pix = doc[0].get_pixmap(dpi=144)             # immutable RGBA8 for NumPy / PIL

# Edit: reorganize pages, merge documents
doc.delete_pages([1, 2])
doc.select([2, 0])                           # keep / reorder / duplicate
merged = pylopdf.Document()
merged.insert_pdf(pylopdf.open("a.pdf"))
merged.insert_pdf(pylopdf.open("b.pdf"), from_page=0, to_page=2)

# Generate: images, overlays, shaped subset-embedded multilingual text
page = doc[0]
page.insert_image((72, 72, 200, 200), filename="logo.png")
page.show_pdf_page(page.rect, letterhead)    # vector overlay from another PDF
page.insert_text((40, 80), "社外秘", fontsize=20, color=(0.8, 0, 0))
page.insert_text((40, 120), "简体中文", fontsize=20, font_language="zh-CN")
page.replace_text("DRAFT", "FINAL")
page.add_highlight_annot(page.search_for("important"))

# Headers / footers / page numbers (standard-14 fonts)
for i, p in enumerate(doc):
    p.insert_text((p.rect.width - 90, p.rect.height - 30), f"Page {i + 1}", fontsize=9)

# Forms (AcroForm)
doc.set_form_field("customer", "Taro Yamada")
doc.set_form_field("agree", True)

# OCR: add a searchable text layer offline (pip install "pylopdf[ocr]")
engine = pylopdf.OcrEngine(threads=4)
page.apply_ocr(engine=engine)                # skips pages with existing text

# Save: optimized, or AES-256 encrypted
doc.save("out.pdf", garbage=True, deflate=True, object_streams=True)
doc.save("locked.pdf", user_pw="secret", permissions=pylopdf.Permissions.PRINT)

# Encrypted input and fast probing
doc = pylopdf.open("locked.pdf", password="secret")
info = pylopdf.peek_metadata("input.pdf")    # page count / encryption, no full parse

The getting started guide walks through each area; the API reference documents every method, default, and resource boundary.

Features

• Editing — merge, split, select/reorder, rotate, page boxes, blank and copied pages, table of contents, page labels, metadata, file attachments, AES-256 encrypted save, atomic file replacement
• Extraction — plain and positioned text (words / blocks / dict), case-insensitive search, bordered and opt-in borderless table detection with merged-cell reconstruction, image and vector-drawing extraction, Markdown conversion, vertical and rotated CJK reading order
• Rendering — PNG and SVG via hayro, page-region clips, DPI control, background fills, ordered parallel render_pages
• Generation — JPEG-passthrough/PNG image insertion, vector page overlays (show_pdf_page), HarfRust-shaped insert_text / insert_textbox with subset embedding and UAX #14 wrapping, reusable Font faces whose text_length is the drawn width, pymupdf's Shape and draw_* vector drawing that embeds each font once per commit, highlight and link annotations, AcroForm filling with regenerated appearances, bounded simple-font text replacement
• OCR — offline PP-OCRv6 engine ([ocr] extra), positioned words, idempotent invisible searchable layers, clockwise rotation correction
• Untrusted input policy — DocumentLimits.web() bounds file size, pages, objects, decompression, glyph and output budgets in one opt-in profile; every documented cap raises a typed LimitError instead of degrading silently (security model)
• WebAssembly — a static PyEmscripten wheel runs on Cloudflare Python Workers, verified end-to-end every release (guide)
• Concurrency — heavy operations release the GIL, distinct Document objects work in parallel, and native cp314t and cp315t wheels support free-threaded Python 3.14 and 3.15 (contract)

Limitations

pylopdf documents its boundaries instead of guessing. Multicolumn reading order follows deterministic whitespace-gutter rules; borderless table detection is opt-in because aligned prose is geometrically ambiguous. CJK text generation selects one font per run (not per-glyph fallback). Pure RTL lines and lines with one Latin or numeric token bracketed by RTL text are restored to logical Unicode order. Edge tokens, multiple tokens, and general mixed bidirectional paragraph layout remain in producer visual order. Ruby, warichu, and mixed-orientation Japanese typography are not interpreted semantically. Typesetting, PDF/A output, and digital signatures are intentionally out of scope — see the ecosystem recipes below. The API stability policy defines what may change before and after v1.0.

Ecosystem recipes

pylopdf stays a lightweight core; adjacent concerns pair with established libraries. These recipes are covered by integration tests and detailed in the ecosystem guide.

Typesetting and PDF/A output — typst (via typst-py):

import typst
import pylopdf

pdf_bytes = typst.compile("report.typ")         # typesetting: typst
doc = pylopdf.open(stream=pdf_bytes)            # editing / extraction: pylopdf
pdf_a: bytes = typst.compile("report.typ", pdf_standards="a-2b")

Digital signatures (PAdES) — pyHanko (MIT), which signs with an incremental update so pylopdf's bytes remain untouched:

import io
from pyhanko.pdf_utils.incremental_writer import IncrementalPdfFileWriter
from pyhanko.sign import signers

signer = signers.SimpleSigner.load("key.pem", "cert.pem")
out = signers.sign_pdf(
    IncrementalPdfFileWriter(io.BytesIO(doc.tobytes())),
    signers.PdfSignatureMetadata(field_name="Signature1"),
    signer=signer,
)

Validating existing PDFs against PDF/A remains veraPDF's job.

Benchmarks

A reproducible benchmark suite ships with the repository — same corpus, same tasks, medians, and wins and losses published as-is. Rendering is reported end to end and split into cold and warm rasterization plus PNG encoding time and size, so a fast PNG encoder cannot hide slower rasterization: latest results, free-threaded extraction, and the native OCR report with strict and NFKC character error rates on licensed Japanese fixtures.

uv sync --all-extras --group bench && uv run python bench/run.py

Architecture

pylopdf.Document (Python, pymupdf-style API)
   └─ _Document (PyO3)
        ├─ lopdf 0.45   … editing: open → modify → save
        ├─ hayro 0.8    … rendering and positioned extraction
        └─ krilla 0.8 + HarfRust 0.14
                         … shaped, subset-embedded text and form appearances

rust/          # PyO3 bindings
src/pylopdf/   # Python high-level API
tests/         # pytest (Rust behavior is verified through Python tests)

Every release installs each published wheel — five abi3-py311 on CPython 3.14 and 3.15, five CPython 3.14t, five CPython 3.15t, the sdist, and the PyEmscripten wheel — on architecture-matched runners and exercises creation, saving, extraction, and rendering before publishing with build provenance.

Contributing

Bug reports and focused contributions are welcome. See CONTRIBUTING.md for development commands, test expectations, and the rules for sharing PDF regression files:

uv sync                    # build + install dependencies
uv run pytest              # tests
uv run ruff check .        # lint
uv run ty check            # type check

Report security vulnerabilities privately through GitHub Security Advisories.

License

MIT. Bundled Rust dependencies keep their own permissive licenses (lopdf and HarfRust are MIT; hayro and krilla are MIT/Apache-2.0) — see NOTICE.md.