yfedoseev/pdf_oxide v0.3.75 on GitHub

Added

ReadingOrder::Structure — order a Tagged PDF by a pre-order traversal of its /StructTreeRoot (ISO 32000-1 §14.8.2.3), fixing table and complex-layout order where a geometric XY-cut guesses; falls back to ColumnAware when the structure tree is absent or not trustworthy (#877).
Mapping provenance — extract_spans reports which ISO 32000-1 §9.10.2 tier (ToUnicode / encoding / heuristic) produced each span's text, surfaced across all bindings (#893).
extract_spans_filtered_with_reading_order — reading-order extraction combined with optional-content (OCG) and ink-coverage filtering (#883).
Type 0 (sampled) and Type 3 (stitching) tint-transform evaluation in the renderer for Separation/DeviceN colours (#849).
CJK vertical-writing running header/footer detection — recognizes folios in the left/right side bands on tategaki pages (#889).
resolve_named_destination is now public (#881); a remove_artifacts markdown-conversion example (#845).

Fixed

Rendering

DeviceCMYK rendered via the naive 1−(C+K) additive clamp — now converts through measured SWOP-style process inks across every composite / vector / text / image path, so 0 0 0 1 k black renders #231F20 and process cyan renders #00ADEF instead of over-saturated additive values (#861).
/Rotate 270 rendered MIRRORED rather than rotated; negative /Rotate values were mishandled (% instead of rem_euclid); real /Rotate and indirect Separation alternates are now resolved correctly (#862, #848, #854).

Text extraction

Form-XObject (Do) text went missing when stray operands preceded the name — a dropped/malformed cm left dangling numeric operands, so Do read operands[0] (a stray number) and resolved to an empty name; it now reads the operand immediately preceding the operator per ISO 32000-1 §7.8.2 (#914).
Pages were lost when /Count-based page counting returned 0 — now recovered by walking the page tree (#909).
Inline images (BI/ID/EI) were parsed but never decoded — /Subtype is now supplied and the Table 92 abbreviated keys/values expanded (#863).
Unique /PlacedPDF bodies are kept instead of suppressed (#896); spans entirely outside the MediaBox are dropped from extract_spans_with_reading_order (#894); the top-level fill colour set before BT is preserved (scn/cs/rg no longer dropped) (#857).
Running-header/footer detection now requires position-consistency before removing repeated text as chrome, and recognizes non-Latin folio digits (#888, #887).

Images

/Decode [1 0] is honoured for 1-bit CCITT/DeviceGray images (#856); the jpeg-decoder Adobe inversion is undone for CMYK JPEGs (#855).

Recovery / parsing

A truncated file that lost its own Catalog is recovered by rebuilding one from the surviving pages (#890); a file padded after %%EOF is no longer rejected outright as 0 pages (#875).

Fonts

The referenced /Encoding /Differences is folded into the font identity hash, and subset choice in extract_embedded_fonts is made deterministic (#878, #853).

Writer

Coloured text on a registered embedded font rendered black — FluentPageBuilder::inline_color (and any TextStyle.color) was silently dropped for embedded fonts — PdfWriter::add_element routed embedded-font text through the deliberately colour-agnostic add_embedded_text (the HTML painter sets and resets the fill colour around its own calls) without ever emitting the element's fill colour, so no rg operator reached the content stream and the glyphs painted in whatever fill colour was last set (default black). The base-14 path (add_text_content) always emits rg from style.color; the embedded path now matches it by emitting fill_color before the glyph run. No restore is needed — every text element sets its own colour, mirroring the base-14 branch's "always set explicitly" contract.

Performance

extract_words / extract_text spent most of their time re-deriving per-glyph facts that cannot change (#882) — text extraction asked each font for its weight and slant once per glyph, inside the show-text loop. Both answers are name-derived: the weight lowercases the base font name (allocating) and runs up to a dozen substring searches, and the slant lowercases it again for two more — so a 13,234-page document repeated ~14 substring scans and 2 allocations 48.7M times for a value fixed at font-load. A sampling profile put str::contains and friends at ~38% of all samples. The Standard-14 width lookup had the same shape, re-stripping the subset prefix and re-scanning the 15-name table per glyph purely to choose a width table. Both are now resolved once per font and memoized, mirroring the existing byte-width-table memo. Alongside: postprocess_spans rescanned every glyph on the page per span to find its baseline (O(spans × chars) — now a bracketed y-sorted index); the page's characters were re-parsed for span post-processing and re-copied on every access (now cached and shared); the word and line paths materialized every glyph twice; article threads — a document-wide parse that walks the whole page tree — were re-parsed per page; the glyph dedup rebuilt the whole array to drop a handful; and the word-merge loop re-derived RTL-ness from an accumulating buffer, costing O(k²) characters per merge chain (the exact blow-up the backtrack guard above it exists to prevent). Measured on the reporter's PDFs: extract_words 156.9s → ~121s and extract_spans 88.7s → ~54s on a 13,234-page document; extract_text 6.03s → 3.94s on a 2,124-page one. Output is byte-identical — verified across a 419-PDF corpus for extract_spans/extract_chars/extract_words/extract_text_lines (including geometry and per-glyph x-offsets) and for text/markdown/HTML.

Changed / Dependencies

Text shaping migrated from rustybuzz to harfrust (#899); fax 0.2 → 0.3 (#873); the ttf-parser migration decision for RUSTSEC-2026-0192 is documented (#900).
office_oxide bumped to 0.1.8 (#904, #932). A combined dependency roundup (crates + CI actions + Go), plus routine crate and CI-action bumps (#931, #907, #898, #872, #869, #870, #871, #864, #865, #866, #867, #868, #874, #891).
Test/CI stability: the flaky structured_warnings round-trip test is fixed (#912); misc test guards and binding-format fixes (#846, #897, #880).

Contributors

The majority of this release was contributed by @ajbufort — rendering (/Rotate handling and Separation/DeviceN tint transforms #848, #849, #854, #862; DeviceCMYK process inks #861), image decode (#855, #856), text extraction (#857, #863, #894, #896, #909), fonts (#853, #878), recovery/parsing (#875, #890), and reading-order / span filtering (#877, #881, #883). Additional fixes from @norbusan (#911) and @ultrasaurus (#845, #846). Thank you!

Issues reported by:

@ankursri494 — #882 (extract_words/extract_text far slower than extract_spans on large PDFs)
@tobocop2 — #876 (signal to callers when a page's text cannot be extracted)
@ultrasaurus — #879 (remove_footers removed real content on IRS forms)
@norbusan — #913 (text inside a Form XObject missing from extraction)

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.