Changelog¶

Release versions come from Git tags (v1.0.0 to v1.0.4 so far). The full history is in CHANGELOG.md and the Git log.

Unreleased¶

Documentation site built from src_docs/md/ with ProperDocs and MaterialX, published at https://vexy.dev/vexy-pdf3md/.
build.sh builds the documentation and the package; publish.sh tags with gitnextver, builds and uploads with uv publish.

Text layers made of control codes (subset fonts without ToUnicode) are detected as scanned pages. Text-layer converters' output of this kind is rejected, and Docling runs in full-page OCR mode for such documents.
--strategies=all runs every strategy side by side.
--language and --vision-languages are replaced by one --lang list (pl,en or pol+eng) used by Tesseract, Docling and pdf22md Vision.
Missing Tesseract models are downloaded from tessdata_fast into ~/.cache/vexy-pdf3md/tessdata. Output configs are mirrored there so Docling's TSV parsing still works with --tessdata-dir. Parallel workers stage downloads in separate files.
Docling: requires pydantic 2.12 or later, and a torchaudio build that fails to load is replaced by a stub.
A document is routed to OCR only when at least 20% of the selected pages look scanned. Isolated image pages log a warning instead.
A failed strategy in a multi-strategy run no longer falls back to a sibling strategy. A warning names the converter whenever it differs from the requested one.
New all extra.

Fire CLI (vexy-pdf3md, pdf3md, python -m vexy_pdf3md) and Python API.
Single-file and recursive-tree conversion, strategy suffixes, and generated media folders with relative links.
Adapters for markit, PyMuPDF, pdf22md, Tesseract and optional Docling.
Scan-aware routing, fallback, text-retention checks, operation timeouts, native-crash isolation, bounded concurrency and per-attempt JSON reports.
Page selection, passwords, OCR controls and languages, DPI and overwrite policy.
Staged publication, collision detection, media ownership checks, and preservation of previous output when publication fails.