turbo-parsepdf

Fast native PDF text / table / image extraction for Python — a pure-Rust core (PyO3, stable-ABI wheels). Imports as turbo_parsepdf. Output as a dict, or HTML / Markdown / JSON strings.

pip install turbo-parsepdf

Benchmark vs the Python PDF stack

Wall-clock to extract every page's text, best-of-N (Apple M-series, release). Reproduce: python benches/competitive-py/bench.py (after python3 benches/gen-corpus.py).

Even including the Python FFI + dict-marshaling overhead, turbo is 38–307× faster — and its text is byte-identical to PyMuPDF (100% word recall).

import turbo_parsepdf

data = open("doc.pdf", "rb").read()

doc = turbo_parsepdf.parse(data)
# {"version": "1.7", "pages": [{"width": ..., "height": ..., "needs_ocr": False,
#   "lines": [{"text": ..., "x": ..., "y": ...}],
#   "tables": [{"rows": ..., "cols": ..., "cells": [[...]]}],
#   "images": [{"name": ..., "format": "Jpeg", "width": ..., ...}]}]}

turbo_parsepdf.parse_to_markdown(data)  # str
turbo_parsepdf.parse_to_html(data)      # str
turbo_parsepdf.parse_to_json(data)      # str

# Encrypted PDFs: pass the user or owner password.
turbo_parsepdf.parse(open("locked.pdf", "rb").read(), password="secret")

A fatal parse fault raises ValueError with a stable code (InvalidHeader, BadStream, …). Scanned/image-only pages come back with needs_ocr=True (OCR is out of scope).

Supports cross-reference streams + object streams (PDF 1.5+), all standard stream filters + predictors, /ToUnicode & encoding/AGL & CID font decoding, ruled tables, image XObject extraction, and standard-handler decryption (RC4 + AES-128/256, R2–R6).

Part of the turbo-parsepdf workspace. MIT.