Skip to content

Releases

pdf-book-to-llm v1.0.0

Choose a tag to compare

@deadboyccc deadboyccc released this 17 Nov 19:53

Folio the bookworm

book2md 1.0.0: turn a PDF book into one Markdown file and one image per page, so you can hand an LLM the exact pages you're asking about. It's installable, robust, and tested against the real Docling models on every push.

uv tool install git+https://github.com/deadboyccc/pdf-book-to-llm
book2md Book.pdf --pages 150-210

What 1.0.0 adds over 0.3.0

  • Folio, the book2md bookworm mascot, and a pipeline picture in the README.
  • README rewritten: features, install, quick start, options and exit codes, LLM usage, troubleshooting, how it is tested. Its output examples are real, unedited output from the CI sample book.
  • Setup guide rewritten for the three install paths (tool, clone, wheel), with the real model download size (~1 GB). New CONTRIBUTING.md, and the decisions are recorded in docs/ASSUMPTIONS.md.
  • A report of the whole effort in reports/.

The road to 1.0

Release Highlights
v0.2.0 Packaging and lockfile, book2md command, --out, exit codes, clean failures, unit tests, CI on Linux, macOS and Windows
v0.3.0 End-to-end tests with the real Docling models on every push

Tested

  • 48 unit tests (99% line coverage) on Linux, macOS and Windows with Python 3.10, 3.12 and 3.13.
  • 10 end-to-end tests with the real Docling models on Linux.
  • All CI jobs green on this commit.

Assets

File What
book2md-1.0.0-py3-none-any.whl Wheel (uv tool install ./book2md-1.0.0-py3-none-any.whl)
book2md-1.0.0.tar.gz Source distribution
SHA256SUMS Checksums

pdf-book-to-llm v0.3.0

Choose a tag to compare

@deadboyccc deadboyccc released this 17 Nov 19:44

book2md is now verified on the real Docling models, on every push. The README's old caveat — that Docling's layout, table and code models had never been exercised — no longer applies.

Added

  • End-to-end tests with real models. CI builds a three-page sample book (headed prose, a bordered table, a code listing, and a raster figure with a caption), runs the installed book2md command on it with the locked Docling version, and checks:
    • a .md and a .png per page, each starting with its <!-- PAGE n --> marker;
    • heading and paragraph text survive;
    • the table becomes a Markdown table with every row;
    • the listing lands in a fenced code block;
    • the figure becomes <!-- image -->, with its caption kept;
    • PNG width follows --scale, and --pages 2-3 writes <book>_p2-3/ with PDF page numbers kept;
    • a start page past the end, or a file that isn't a PDF, fails with one line and writes nothing.
  • The converted sample pages are uploaded as a CI artifact, so you can see real output for each commit.
  • CI also runs on milestone-* branches, so heavy changes go green before main moves.

Observed in real output

  • Code blocks are detected and fenced, but indentation may come back as tabs, and no language tag is set.

Tested

  • 48 unit tests on Linux, macOS and Windows, plus 10 end-to-end tests with real Docling models (about 50 s in CI with cached weights).

Assets

File What
book2md-0.3.0-py3-none-any.whl Wheel
book2md-0.3.0.tar.gz Source distribution
SHA256SUMS Checksums

pdf-book-to-llm v0.2.0

Choose a tag to compare

@deadboyccc deadboyccc released this 17 Nov 19:39

First release of book2md: the converter is now an installable, tested project. It is reproducible from its lockfile and robust against the mistakes users actually make.

Added

  • Installable command: uv tool install git+https://github.com/deadboyccc/pdf-book-to-llm, then run book2md. uv run python book2md.py still works.
  • pyproject.toml + uv.lock, so uv sync from a clone works as the setup guide always said. Docling 2.150 is locked.
  • --out DIR to choose the output root (default ~/study/corpus); previously you had to edit the script. Also --version.
  • Exit codes: 0 all pages written, 1 failed or partial, 2 invalid arguments.
  • CI: lint, lockfile check, unit tests on Linux, macOS and Windows (Python 3.10, 3.12, 3.13), and a package build.

Fixed

  • A start page past the end of the book, or a file that isn't a PDF, printed a Docling traceback and left an empty output folder behind. Now it prints one error line and writes nothing.
  • A partial conversion silently wrote fewer pages and printed "Done". Now the missing pages are listed with Docling's errors, and the exit code is 1.
  • --scale 0 or a negative scale only failed deep inside Docling. Now it's a usage error.
  • Both docs linked to files that did not exist (docs/setup-and-script.md, README.md).
  • --help waited seconds for Docling to import. Docling now loads only when converting.

Tested

  • 0 → 48 unit tests, with Docling replaced by fakes; 99% line coverage.
  • Validated from a fresh clone in a sandbox; CI green on every job.

Assets

File What
book2md-0.2.0-py3-none-any.whl Wheel (uv tool install ./book2md-0.2.0-py3-none-any.whl)
book2md-0.2.0.tar.gz Source distribution
SHA256SUMS Checksums