v0.3.54 | Text-extraction fidelity pass. Hebrew / RTL reads in correct reading order thanks to a new geometric visual-vs-logical detector that finally fires on Hebrew (Pass 0's Arabic-Presentation-Forms gate never matched Hebrew because Hebrew has no presentation forms in Unicode). Bullets and `fi` / `fl` ligatures decode to their canonical codepoints via a new ISO 32000-1 §9.10.2 fallback level — Type0 Identity-H fonts without a `CIDToGIDMap` now consult the embedded TrueType `cmap` + `post` tables before the CID-as-Unicode last resort. Tight two-column prose bodies read column-by-column instead of row-by-row interleave (the known v0.3.53 issue), gated by a new table-vs-prose classifier so the v0.3.53 corpus-revert lesson holds — tables stay byte-identical. Two-column reference PDFs (Bibles / dictionaries / encyclopedias) inherit the same reading-order fix. Four fixes; zero new public surface.

Fixed

• Hebrew / RTL visual-vs-logical detection
  (#537)
  — Hebrew PDFs that store text in visual order (the PDF
  content stream draws glyphs left-to-right even though the
  script reads right-to-left) now extract in correct logical
  order. New per-RTL-run X-coordinate-monotonicity detector
  gates the existing UAX #9 bidi::reorder_visual_to_logical
  pass; logical-order PDFs (the pdfium hebrew_mirrored.pdf
  test fixture and Arabic CID-TrueType samples) stay
  byte-identical. Pass 0's Arabic-Presentation-Forms trigger
  was the only Arabic-specific gate; Hebrew has no
  presentation forms in Unicode, so the geometric gate is
  what unblocks it. Reported by alexagr with the Magic
  Palace Eilat hotel PDF as the canonical case; reproduced
  empirically before the fix landed.

• ToUnicode CMap miss: extended ISO 32000-1 §9.10.2
  fallback chain
  (#535)
  — Type0 Identity-H fonts without a CIDToGIDMap (the
  PowerPoint / Acrobat-exported subset-font shape that
  produced ~25,350 "ToUnicode CMap MISS" warnings on the
  PowerPoint corpus) now consult the embedded TrueType
  font's cmap + post tables to recover glyph names,
  then look up the Adobe Glyph List (including the AGL §6
  uniXXXX / uXXXXX synthetic-name patterns). Resolves
  the bullet → ❍ substitution and the fi / fl
  ligature corruption. The post-extraction band-aid
  normalize_bullet_glyphs stays retired (it was wrong for
  documents that legitimately use ❍); the fix is in the
  decoder, where it belongs. Same shape pdf.js / MuPDF /
  PDFBox 3.x use.

• Multi-column prose reading order
  (#534)
  — tight two-column PROSE bodies (~12pt gutters) now read
  column-by-column instead of row-by-row interleave. The
  v0.3.53 known issue, explicitly documented in that
  release's notes. New table-vs-prose classifier (region
  is "prose" when lines are tall stacks of wide content,
  "table" when lines are short cells; the v0.3.53 revert
  comment in xycut.rs:73-101 is the spec) gates a
  tight-gutter column cut so the same XY-cut recursion no
  longer corrupts table cells — the v0.3.53 lesson where two
  prior attempts were reverted by the 70-PDF corpus sweep
  after the google_doc_document.pdf population table
  ("273.879.7501" → "1273.879.750") regressed. Adds
  recursive band-separation of full-width header / footer
  rows BEFORE column detection so the multi-column body
  gets analyzed as its own region. Same 70-PDF sweep is the
  acceptance gate.

• Two-column reference PDFs: column-aware reading order
  (#536)
  — reference-style two-column PDFs (Bibles, dictionaries,
  encyclopedias, academic editions with marginal section
  numbers) now extract column-by-column thanks to the #534
  classifier. The cascade where the spatial-table-detector
  latched onto the interleaved structure and rendered the
  whole multi-column body as a Markdown table-with-
  one-word-per-cell is resolved: clean column reading order
  produces clean prose, not a degenerate table. Concerns
  about marginal-label classification, structural heading
  isolation, cross-page section continuity, and
  header / footer separation that the reporter usefully
  called out are captured for the v0.4.0
  structured-extraction track and acknowledged on the
  issue. Reported by lggcs (Luke) with a 1256-page
  French Louis Segond Bible as the canonical case;
  reproduced empirically before the fix landed.

---

Installation

Rust (crates.io)

cargo add pdf_oxide

Python (PyPI)

pip install pdf_oxide

JavaScript/WASM (npm)

npm install pdf-oxide-wasm

CLI (Homebrew)

brew install yfedoseev/tap/pdf-oxide

CLI (Scoop — Windows)

scoop bucket add pdf-oxide https://github.com/yfedoseev/scoop-pdf-oxide
scoop install pdf-oxide

CLI (Shell installer)

curl -fsSL https://raw.githubusercontent.com/yfedoseev/pdf_oxide/main/install.sh | sh

CLI (cargo-binstall)

cargo binstall pdf_oxide_cli

MCP Server (for AI assistants)

cargo install pdf_oxide_mcp

Pre-built Binaries
Download archives for Linux, macOS, and Windows from the assets below. Each archive includes both pdf-oxide (CLI) and pdf-oxide-mcp (MCP server).

Platform Support

Platform	Architecture	Archive
Linux	x86_64 (glibc)	pdf_oxide-linux-x86_64-*.tar.gz
Linux	x86_64 (musl)	pdf_oxide-linux-x86_64-musl-*.tar.gz
Linux	ARM64	pdf_oxide-linux-aarch64-*.tar.gz
macOS	x86_64 (Intel)	pdf_oxide-macos-x86_64-*.tar.gz
macOS	ARM64 (Apple Silicon)	pdf_oxide-macos-aarch64-*.tar.gz
Windows	x86_64	pdf_oxide-windows-x86_64-*.zip

Changelog

See CHANGELOG.md for full details.