TL;DR
zpdf is an alpha-stage PDF text extraction library implemented in Zig that emphasizes speed through memory mapping, SIMD-accelerated routines and multi-threaded page extraction. Benchmarks in the repository show large speedups versus MuPDF for text extraction on Apple M4 Pro and Intel test data, while the project notes limitations for complex PDF features and non-Latin scripts.
What happened
A new open-source library named zpdf provides a Zig-based implementation for extracting text from PDF files. The project targets high throughput for large and batch workloads by combining memory-mapped file access, streaming extraction with arena allocation, multi-threaded page processing and SIMD-optimized string routines (NEON, AVX2/SSE4.2 or scalar fallback). The repository includes a command-line tool and Python bindings (via cffi), usage examples, and instructions to build with Zig. Benchmarks published in the source compare zpdf to MuPDF’s text-extraction path, reporting multi-fold speedups in sequential and parallel modes and a reported peak throughput of 45,000 pages/sec on an Intel document test. Accuracy tests against MuPDF on one sample (US Constitution) report 99.6% character similarity and a 2.1% word error rate for zpdf. The codebase is described as alpha and the feature matrix lists missing support for embedded fonts, certain image codecs and encrypted PDFs.
Why it matters
- Faster text extraction can drastically reduce processing time in bulk workflows such as large-scale indexing or data pipelines.
- Parallel page extraction and SIMD acceleration promise better utilization of modern multi-core and SIMD-capable CPUs for I/O-bound PDF jobs.
- Lightweight, language-bindable toolchain (Zig + C ABI + Python bindings) makes integration into diverse tool stacks possible without shipping a full renderer.
- Limitations around embedded fonts, some CMap handling and non-Latin scripts mean zpdf may not replace full-featured PDF engines for complex documents.
Key facts
- Project status: alpha / early version.
- Implemented in Zig; requires Zig 0.15.2 or later to build.
- Performance: repository benchmarks show 3.9x–4.7x sequential speedups versus MuPDF on several documents and larger parallel speedups (up to 17.9x on one Intel data set).
- Peak throughput claimed: 45,000 pages/sec (Intel SDM, parallel test).
- SIMD-accelerated hot paths include whitespace skipping, delimiter detection, keyword search and string boundary scanning; auto-detects NEON, AVX2/SSE4.2 or falls back to scalar code.
- Supported decompression filters include FlateDecode, ASCII85, ASCIIHex, LZW and RunLength.
- Font encoding support: WinAnsi, MacRoman and ToUnicode CMap (with caveats for compressed object streams).
- Missing or partial support: embedded fonts, JBIG2, JPEG2000, encrypted PDFs, forms/annotations and rendering.
- Provided interfaces: Zig library API, a CLI with page selection and an experimental reading-order mode, and Python bindings via cffi.
- License: WTFPL.
What to watch next
- The project’s experimental reading-order extraction feature — currently labeled experimental in the source.
- Improvements to ToUnicode/CID font handling where references to compressed object streams are involved (source notes partial support).
- Not confirmed in the source: a roadmap or timeline for adding embedded font support, JBIG2/JPEG2000 codecs, encrypted PDF handling or full CID font coverage.
- Not confirmed in the source: plans for a stable release or long-term maintenance commitments.
Quick glossary
- SIMD: Single Instruction, Multiple Data — a CPU feature that lets one instruction operate on multiple data points at once, useful for accelerating string and numeric operations.
- XRef table/stream: A PDF structure that maps object numbers to locations in the file; parsing it is necessary to locate and read objects within a PDF.
- ToUnicode CMap: A mapping inside a PDF that translates character codes used by a font into Unicode code points for accurate text extraction.
- WER: Word Error Rate — a measure of text extraction or recognition errors at the word level, lower values indicate fewer errors.
Reader FAQ
Is zpdf production-ready?
The repository describes zpdf as alpha / early stage; users should treat it as experimental.
How do I build and run the project?
The source includes build commands: 'zig build' to build the library/CLI and 'zig build test' to run tests; the benchmarks were produced with 'zig build -Doptimize=ReleaseFast'.
Does zpdf support encrypted PDFs?
Not supported in zpdf according to the project's feature comparison.
Are there Python bindings?
Yes — the repository includes Python bindings implemented via cffi and example usage.
Is reading-order extraction available?
A reading-order extraction mode exists but is marked experimental in the source.
zpdf (alpha stage – early version) A PDF text extraction library written in Zig. Features Memory-mapped file reading for efficient large file handling Streaming text extraction with efficient arena allocation…
Sources
- Zpdf: PDF text extraction in Zig
- Zpdf: PDF text extraction in Zig – 5x faster than MuPDF
- zpdf/build.zig at main
- Awesome Zig
Related posts
- S&P 500 Priced in Gold: 1880–2012 Data Reveal Large Real Decline
- F-Droid replaces core server with donated hardware to speed builds
- 73 Programming Project Ideas to Inspire and Challenge Developers Worldwide