pdf-rtldoc 1.0.6__tar.gz

pdf_rtldoc-1.0.6/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Umer Javaid
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

pdf_rtldoc-1.0.6/PKG-INFO

@@ -0,0 +1,163 @@
+Metadata-Version: 2.4
+Name: pdf-rtldoc
+Version: 1.0.6
+Summary: Geometry-first PDF parser for RTL and complex layouts (Arabic textbooks, financial tables, multi-column docs)
+License: MIT
+Project-URL: Homepage, https://github.com/umerjavaidkh/rtldoc
+Project-URL: Repository, https://github.com/umerjavaidkh/rtldoc
+Project-URL: Issues, https://github.com/umerjavaidkh/rtldoc/issues
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: pymupdf>=1.24
+Requires-Dist: numpy
+Provides-Extra: cv
+Requires-Dist: opencv-python-headless; extra == "cv"
+Provides-Extra: dev
+Requires-Dist: arabic-reshaper; extra == "dev"
+Requires-Dist: python-bidi; extra == "dev"
+Dynamic: license-file
+
+# rtldoc
+
+**A geometry-first PDF parser for RTL & complex layouts.** It extracts text,
+reading order, and tables from born-digital PDFs using the glyph coordinates
+already in the file — **no OCR, no GPU, no API**. Built for the pages that
+quietly break Docling, Marker, and VLM parsers: Arabic/RTL, multi-column,
+and vector tables.
+
+---
+
+## Why it's better where it matters
+
+- **RTL / Arabic done right.** Reading order and bidi are rebuilt from glyph
+  *positions*, not from a reading-order model trained on English. It fixes the
+  presentation-form and lam-alef bugs that silently corrupt ~⅓ of Arabic words
+  in every general parser — encoding bugs no layout model can fine-tune away.
+- **Tables without a model.** Reads table structure from the PDF's own vector
+  rules at 100% (vs ~85–95% from a CNN), and recovers *borderless* tables by
+  column alignment.
+- **Deterministic & auditable.** Every output traces to a rule you can point
+  at. CPU-only, **6–60 pages/sec**, $0/page — a VLM is 100–1000× the cost and
+  can't be audited.
+- **Page visual summary, no vision model.** Opt-in (`--visual`) geometry pass
+  that turns a page's own vector drawings into a structured summary: image
+  size/color stats, table dimensions, and — for vector flowcharts/diagrams —
+  the actual boxes, their text labels, and which ones a connecting line joins.
+  Built entirely from the PDF's own drawing commands (PyMuPDF `get_drawings()`),
+  not a screenshot or a guess. HTML output renders any detected diagram as a
+  real chart via Mermaid.js, fed the extracted nodes/edges directly.
+
+## Proven at scale
+
+Tested on **119 PDFs / 13,557 pages** it never saw during development — an
+Arabic teacher's guide, two SEC 10-Ks, 96+ arXiv papers (15 fields), 5 OpenStax
+physics/chemistry/calculus textbooks (figures, geometry, exercises), the
+3,130-page PostgreSQL 18 manual (deeply-nested reference tables, code blocks),
+and a growing set of real-world forms, reports, and scanned documents.
+The checks are *property-based and label-free*, so they scale to any corpus:
+
+| Property (must hold on every page) | Result |
+|---|---:|
+| crashes | **0** |
+| encoding leaks (presentation forms in output) | **0** |
+| malformed tables | **0** |
+| non-deterministic pages | **0** |
+| text coverage vs the PDF's own glyph stream | **~99%** |
+
+**Table quality, scored with TEDS** (the PubTabNet/OmniDocBench standard) on a
+borderless financial statement — where the whole point is a hard table:
+
+| rtldoc | pdfplumber | naive `get_text` |
+|---:|---:|---:|
+| **0.942** | 0.061 | 0.000 |
+
+Everything is reproducible in [`eval/`](eval/) (harnesses, arXiv manifest,
+saved reports).
+
+## Use it
+
+```bash
+pip install pdf-rtldoc
+```
+
+(Published on PyPI as `pdf-rtldoc` -- the plain `rtldoc` name was already
+taken by an unrelated project. The installed command, and the module you
+import in Python, are both still `rtldoc`.)
+
+Or install straight from a specific release without going through PyPI:
+
+```bash
+pip install "git+https://github.com/umerjavaidkh/rtldoc.git@v1.0.6"
+```
+
+Then:
+
+```bash
+rtldoc parse book.pdf --md out/ --json out.json
+rtldoc parse book.pdf --html out_html/   # real <table>/<figure>, RTL-aware dir=
+rtldoc parse book.pdf --html out_html/ --visual  # + diagram/image/table visual summary
+rtldoc audit book.pdf                    # flags low-confidence pages for review
+```
+
+Runtime deps are just PyMuPDF + numpy. (From a clone: `pip install -e .`.)
+
+Zero-setup via Docker (399 MB, no compiler/GPU):
+
+```bash
+docker build -t rtldoc . && docker run --rm -v "$PWD:/d" rtldoc parse /d/book.pdf --md /d/out
+```
+
+Output: per-page Markdown (tables as GFM, images extracted + auto-captioned)
+plus structured JSON, or a self-contained HTML page per PDF page.
+
+## Honest limits
+
+- Borderless-table *grid geometry* is approximate on the hardest wide,
+  multi-level-header tables (occasional row/column structure mismatches
+  — that's the 0.942 TEDS on our graded case, not 1.0).
+- A page dominated by a figure/table spanning the full content width can
+  still be under-counted as fewer columns than it visually has — the
+  whitespace-gutter detector requires the gap to stay empty across most
+  of the page's height, and a full-width element defeats that locally.
+- A table cell whose own text wraps onto a later line, where that line's
+  content coincidentally re-aligns with an earlier row's columns, can
+  occasionally attach to the wrong row (`_merge_wrapped_label_rows` — a
+  narrow, tracked edge case, not a general table-detection failure).
+- Scanned / no-text-layer pages now OCR via Tesseract (needs the `tesseract`
+  binary on PATH — `brew install tesseract` / `apt install tesseract-ocr`;
+  no Python package required). Word-level positioning, not this repo's
+  glyph-exact reading order; a page with no tesseract installed just gets
+  no blocks, as before.
+- Best semantic typing needs a one-time per-publisher style map (~20 min).
+- Diagram detection reconstructs simple box-and-arrow flowcharts reliably;
+  dense multi-level diagrams (deep tree/org-chart hierarchies with many
+  branches) get correct node/box detection but not yet reliable connection
+  tracing — a harder, separate problem noted for future work.
+
+## Roadmap
+
+- Reliable connection tracing for dense/branching diagrams (see above).
+- A page-level chart/figure classification pass, so bar charts, legends,
+  and gridlines are recognized and set aside before table/diagram
+  detection runs, rather than relying on those detectors' own guards to
+  reject them case by case.
+- A cell-level golden regression corpus (`eval/golden/` + `eval/regression.py`)
+  now exists and grows with each table-detection fix; still short of full
+  coverage across document types.
+
+## Publishing a release (maintainers)
+
+Bump `version` in `pyproject.toml`, commit, tag (`git tag -a vX.Y.Z`), push
+the tag, then publish a GitHub Release from it. `.github/workflows/publish.yml`
+builds and uploads to PyPI automatically when the release is published, via
+PyPI's Trusted Publisher (OIDC) mechanism — no API token stored anywhere.
+
+One-time setup (already done for `pdf-rtldoc`): on pypi.org, under the
+project's *Publishing* settings, add a trusted publisher with owner
+`umerjavaidkh`, repository `rtldoc`, workflow filename `publish.yml`, and
+environment name `pypi`.
+
+---
+
+*Design rationale, the full bug log, and methodology: **[DESIGN.md](DESIGN.md)**.*

pdf_rtldoc-1.0.6/README.md

@@ -0,0 +1,143 @@
+# rtldoc
+
+**A geometry-first PDF parser for RTL & complex layouts.** It extracts text,
+reading order, and tables from born-digital PDFs using the glyph coordinates
+already in the file — **no OCR, no GPU, no API**. Built for the pages that
+quietly break Docling, Marker, and VLM parsers: Arabic/RTL, multi-column,
+and vector tables.
+
+---
+
+## Why it's better where it matters
+
+- **RTL / Arabic done right.** Reading order and bidi are rebuilt from glyph
+  *positions*, not from a reading-order model trained on English. It fixes the
+  presentation-form and lam-alef bugs that silently corrupt ~⅓ of Arabic words
+  in every general parser — encoding bugs no layout model can fine-tune away.
+- **Tables without a model.** Reads table structure from the PDF's own vector
+  rules at 100% (vs ~85–95% from a CNN), and recovers *borderless* tables by
+  column alignment.
+- **Deterministic & auditable.** Every output traces to a rule you can point
+  at. CPU-only, **6–60 pages/sec**, $0/page — a VLM is 100–1000× the cost and
+  can't be audited.
+- **Page visual summary, no vision model.** Opt-in (`--visual`) geometry pass
+  that turns a page's own vector drawings into a structured summary: image
+  size/color stats, table dimensions, and — for vector flowcharts/diagrams —
+  the actual boxes, their text labels, and which ones a connecting line joins.
+  Built entirely from the PDF's own drawing commands (PyMuPDF `get_drawings()`),
+  not a screenshot or a guess. HTML output renders any detected diagram as a
+  real chart via Mermaid.js, fed the extracted nodes/edges directly.
+
+## Proven at scale
+
+Tested on **119 PDFs / 13,557 pages** it never saw during development — an
+Arabic teacher's guide, two SEC 10-Ks, 96+ arXiv papers (15 fields), 5 OpenStax
+physics/chemistry/calculus textbooks (figures, geometry, exercises), the
+3,130-page PostgreSQL 18 manual (deeply-nested reference tables, code blocks),
+and a growing set of real-world forms, reports, and scanned documents.
+The checks are *property-based and label-free*, so they scale to any corpus:
+
+| Property (must hold on every page) | Result |
+|---|---:|
+| crashes | **0** |
+| encoding leaks (presentation forms in output) | **0** |
+| malformed tables | **0** |
+| non-deterministic pages | **0** |
+| text coverage vs the PDF's own glyph stream | **~99%** |
+
+**Table quality, scored with TEDS** (the PubTabNet/OmniDocBench standard) on a
+borderless financial statement — where the whole point is a hard table:
+
+| rtldoc | pdfplumber | naive `get_text` |
+|---:|---:|---:|
+| **0.942** | 0.061 | 0.000 |
+
+Everything is reproducible in [`eval/`](eval/) (harnesses, arXiv manifest,
+saved reports).
+
+## Use it
+
+```bash
+pip install pdf-rtldoc
+```
+
+(Published on PyPI as `pdf-rtldoc` -- the plain `rtldoc` name was already
+taken by an unrelated project. The installed command, and the module you
+import in Python, are both still `rtldoc`.)
+
+Or install straight from a specific release without going through PyPI:
+
+```bash
+pip install "git+https://github.com/umerjavaidkh/rtldoc.git@v1.0.6"
+```
+
+Then:
+
+```bash
+rtldoc parse book.pdf --md out/ --json out.json
+rtldoc parse book.pdf --html out_html/   # real <table>/<figure>, RTL-aware dir=
+rtldoc parse book.pdf --html out_html/ --visual  # + diagram/image/table visual summary
+rtldoc audit book.pdf                    # flags low-confidence pages for review
+```
+
+Runtime deps are just PyMuPDF + numpy. (From a clone: `pip install -e .`.)
+
+Zero-setup via Docker (399 MB, no compiler/GPU):
+
+```bash
+docker build -t rtldoc . && docker run --rm -v "$PWD:/d" rtldoc parse /d/book.pdf --md /d/out
+```
+
+Output: per-page Markdown (tables as GFM, images extracted + auto-captioned)
+plus structured JSON, or a self-contained HTML page per PDF page.
+
+## Honest limits
+
+- Borderless-table *grid geometry* is approximate on the hardest wide,
+  multi-level-header tables (occasional row/column structure mismatches
+  — that's the 0.942 TEDS on our graded case, not 1.0).
+- A page dominated by a figure/table spanning the full content width can
+  still be under-counted as fewer columns than it visually has — the
+  whitespace-gutter detector requires the gap to stay empty across most
+  of the page's height, and a full-width element defeats that locally.
+- A table cell whose own text wraps onto a later line, where that line's
+  content coincidentally re-aligns with an earlier row's columns, can
+  occasionally attach to the wrong row (`_merge_wrapped_label_rows` — a
+  narrow, tracked edge case, not a general table-detection failure).
+- Scanned / no-text-layer pages now OCR via Tesseract (needs the `tesseract`
+  binary on PATH — `brew install tesseract` / `apt install tesseract-ocr`;
+  no Python package required). Word-level positioning, not this repo's
+  glyph-exact reading order; a page with no tesseract installed just gets
+  no blocks, as before.
+- Best semantic typing needs a one-time per-publisher style map (~20 min).
+- Diagram detection reconstructs simple box-and-arrow flowcharts reliably;
+  dense multi-level diagrams (deep tree/org-chart hierarchies with many
+  branches) get correct node/box detection but not yet reliable connection
+  tracing — a harder, separate problem noted for future work.
+
+## Roadmap
+
+- Reliable connection tracing for dense/branching diagrams (see above).
+- A page-level chart/figure classification pass, so bar charts, legends,
+  and gridlines are recognized and set aside before table/diagram
+  detection runs, rather than relying on those detectors' own guards to
+  reject them case by case.
+- A cell-level golden regression corpus (`eval/golden/` + `eval/regression.py`)
+  now exists and grows with each table-detection fix; still short of full
+  coverage across document types.
+
+## Publishing a release (maintainers)
+
+Bump `version` in `pyproject.toml`, commit, tag (`git tag -a vX.Y.Z`), push
+the tag, then publish a GitHub Release from it. `.github/workflows/publish.yml`
+builds and uploads to PyPI automatically when the release is published, via
+PyPI's Trusted Publisher (OIDC) mechanism — no API token stored anywhere.
+
+One-time setup (already done for `pdf-rtldoc`): on pypi.org, under the
+project's *Publishing* settings, add a trusted publisher with owner
+`umerjavaidkh`, repository `rtldoc`, workflow filename `publish.yml`, and
+environment name `pypi`.
+
+---
+
+*Design rationale, the full bug log, and methodology: **[DESIGN.md](DESIGN.md)**.*

pdf_rtldoc-1.0.6/pdf_rtldoc.egg-info/PKG-INFO

@@ -0,0 +1,163 @@
+Metadata-Version: 2.4
+Name: pdf-rtldoc
+Version: 1.0.6
+Summary: Geometry-first PDF parser for RTL and complex layouts (Arabic textbooks, financial tables, multi-column docs)
+License: MIT
+Project-URL: Homepage, https://github.com/umerjavaidkh/rtldoc
+Project-URL: Repository, https://github.com/umerjavaidkh/rtldoc
+Project-URL: Issues, https://github.com/umerjavaidkh/rtldoc/issues
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: pymupdf>=1.24
+Requires-Dist: numpy
+Provides-Extra: cv
+Requires-Dist: opencv-python-headless; extra == "cv"
+Provides-Extra: dev
+Requires-Dist: arabic-reshaper; extra == "dev"
+Requires-Dist: python-bidi; extra == "dev"
+Dynamic: license-file
+
+# rtldoc
+
+**A geometry-first PDF parser for RTL & complex layouts.** It extracts text,
+reading order, and tables from born-digital PDFs using the glyph coordinates
+already in the file — **no OCR, no GPU, no API**. Built for the pages that
+quietly break Docling, Marker, and VLM parsers: Arabic/RTL, multi-column,
+and vector tables.
+
+---
+
+## Why it's better where it matters
+
+- **RTL / Arabic done right.** Reading order and bidi are rebuilt from glyph
+  *positions*, not from a reading-order model trained on English. It fixes the
+  presentation-form and lam-alef bugs that silently corrupt ~⅓ of Arabic words
+  in every general parser — encoding bugs no layout model can fine-tune away.
+- **Tables without a model.** Reads table structure from the PDF's own vector
+  rules at 100% (vs ~85–95% from a CNN), and recovers *borderless* tables by
+  column alignment.
+- **Deterministic & auditable.** Every output traces to a rule you can point
+  at. CPU-only, **6–60 pages/sec**, $0/page — a VLM is 100–1000× the cost and
+  can't be audited.
+- **Page visual summary, no vision model.** Opt-in (`--visual`) geometry pass
+  that turns a page's own vector drawings into a structured summary: image
+  size/color stats, table dimensions, and — for vector flowcharts/diagrams —
+  the actual boxes, their text labels, and which ones a connecting line joins.
+  Built entirely from the PDF's own drawing commands (PyMuPDF `get_drawings()`),
+  not a screenshot or a guess. HTML output renders any detected diagram as a
+  real chart via Mermaid.js, fed the extracted nodes/edges directly.
+
+## Proven at scale
+
+Tested on **119 PDFs / 13,557 pages** it never saw during development — an
+Arabic teacher's guide, two SEC 10-Ks, 96+ arXiv papers (15 fields), 5 OpenStax
+physics/chemistry/calculus textbooks (figures, geometry, exercises), the
+3,130-page PostgreSQL 18 manual (deeply-nested reference tables, code blocks),
+and a growing set of real-world forms, reports, and scanned documents.
+The checks are *property-based and label-free*, so they scale to any corpus:
+
+| Property (must hold on every page) | Result |
+|---|---:|
+| crashes | **0** |
+| encoding leaks (presentation forms in output) | **0** |
+| malformed tables | **0** |
+| non-deterministic pages | **0** |
+| text coverage vs the PDF's own glyph stream | **~99%** |
+
+**Table quality, scored with TEDS** (the PubTabNet/OmniDocBench standard) on a
+borderless financial statement — where the whole point is a hard table:
+
+| rtldoc | pdfplumber | naive `get_text` |
+|---:|---:|---:|
+| **0.942** | 0.061 | 0.000 |
+
+Everything is reproducible in [`eval/`](eval/) (harnesses, arXiv manifest,
+saved reports).
+
+## Use it
+
+```bash
+pip install pdf-rtldoc
+```
+
+(Published on PyPI as `pdf-rtldoc` -- the plain `rtldoc` name was already
+taken by an unrelated project. The installed command, and the module you
+import in Python, are both still `rtldoc`.)
+
+Or install straight from a specific release without going through PyPI:
+
+```bash
+pip install "git+https://github.com/umerjavaidkh/rtldoc.git@v1.0.6"
+```
+
+Then:
+
+```bash
+rtldoc parse book.pdf --md out/ --json out.json
+rtldoc parse book.pdf --html out_html/   # real <table>/<figure>, RTL-aware dir=
+rtldoc parse book.pdf --html out_html/ --visual  # + diagram/image/table visual summary
+rtldoc audit book.pdf                    # flags low-confidence pages for review
+```
+
+Runtime deps are just PyMuPDF + numpy. (From a clone: `pip install -e .`.)
+
+Zero-setup via Docker (399 MB, no compiler/GPU):
+
+```bash
+docker build -t rtldoc . && docker run --rm -v "$PWD:/d" rtldoc parse /d/book.pdf --md /d/out
+```
+
+Output: per-page Markdown (tables as GFM, images extracted + auto-captioned)
+plus structured JSON, or a self-contained HTML page per PDF page.
+
+## Honest limits
+
+- Borderless-table *grid geometry* is approximate on the hardest wide,
+  multi-level-header tables (occasional row/column structure mismatches
+  — that's the 0.942 TEDS on our graded case, not 1.0).
+- A page dominated by a figure/table spanning the full content width can
+  still be under-counted as fewer columns than it visually has — the
+  whitespace-gutter detector requires the gap to stay empty across most
+  of the page's height, and a full-width element defeats that locally.
+- A table cell whose own text wraps onto a later line, where that line's
+  content coincidentally re-aligns with an earlier row's columns, can
+  occasionally attach to the wrong row (`_merge_wrapped_label_rows` — a
+  narrow, tracked edge case, not a general table-detection failure).
+- Scanned / no-text-layer pages now OCR via Tesseract (needs the `tesseract`
+  binary on PATH — `brew install tesseract` / `apt install tesseract-ocr`;
+  no Python package required). Word-level positioning, not this repo's
+  glyph-exact reading order; a page with no tesseract installed just gets
+  no blocks, as before.
+- Best semantic typing needs a one-time per-publisher style map (~20 min).
+- Diagram detection reconstructs simple box-and-arrow flowcharts reliably;
+  dense multi-level diagrams (deep tree/org-chart hierarchies with many
+  branches) get correct node/box detection but not yet reliable connection
+  tracing — a harder, separate problem noted for future work.
+
+## Roadmap
+
+- Reliable connection tracing for dense/branching diagrams (see above).
+- A page-level chart/figure classification pass, so bar charts, legends,
+  and gridlines are recognized and set aside before table/diagram
+  detection runs, rather than relying on those detectors' own guards to
+  reject them case by case.
+- A cell-level golden regression corpus (`eval/golden/` + `eval/regression.py`)
+  now exists and grows with each table-detection fix; still short of full
+  coverage across document types.
+
+## Publishing a release (maintainers)
+
+Bump `version` in `pyproject.toml`, commit, tag (`git tag -a vX.Y.Z`), push
+the tag, then publish a GitHub Release from it. `.github/workflows/publish.yml`
+builds and uploads to PyPI automatically when the release is published, via
+PyPI's Trusted Publisher (OIDC) mechanism — no API token stored anywhere.
+
+One-time setup (already done for `pdf-rtldoc`): on pypi.org, under the
+project's *Publishing* settings, add a trusted publisher with owner
+`umerjavaidkh`, repository `rtldoc`, workflow filename `publish.yml`, and
+environment name `pypi`.
+
+---
+
+*Design rationale, the full bug log, and methodology: **[DESIGN.md](DESIGN.md)**.*

pdf_rtldoc-1.0.6/pdf_rtldoc.egg-info/SOURCES.txt

@@ -0,0 +1,19 @@
+LICENSE
+README.md
+pyproject.toml
+pdf_rtldoc.egg-info/PKG-INFO
+pdf_rtldoc.egg-info/SOURCES.txt
+pdf_rtldoc.egg-info/dependency_links.txt
+pdf_rtldoc.egg-info/entry_points.txt
+pdf_rtldoc.egg-info/requires.txt
+pdf_rtldoc.egg-info/top_level.txt
+rtldoc/__init__.py
+rtldoc/arabic.py
+rtldoc/cli.py
+rtldoc/cvfallback.py
+rtldoc/geobidi.py
+rtldoc/layout.py
+rtldoc/ocr.py
+rtldoc/pipeline.py
+rtldoc/primitives.py
+rtldoc/visual.py

pdf_rtldoc-1.0.6/pdf_rtldoc.egg-info/dependency_links.txt

@@ -0,0 +1 @@
+

pdf_rtldoc-1.0.6/pdf_rtldoc.egg-info/entry_points.txt

@@ -0,0 +1,2 @@
+[console_scripts]
+rtldoc = rtldoc.cli:main

pdf_rtldoc-1.0.6/pdf_rtldoc.egg-info/requires.txt

@@ -0,0 +1,9 @@
+pymupdf>=1.24
+numpy
+
+[cv]
+opencv-python-headless
+
+[dev]
+arabic-reshaper
+python-bidi

pdf_rtldoc-1.0.6/pdf_rtldoc.egg-info/top_level.txt

@@ -0,0 +1 @@
+rtldoc

pdf_rtldoc-1.0.6/pyproject.toml

@@ -0,0 +1,44 @@
+[build-system]
+requires = ["setuptools>=68"]
+build-backend = "setuptools.build_meta"
+
+[project]
+# PyPI distribution name is "pdf-rtldoc" -- "rtldoc" was already taken by
+# an unrelated SystemVerilog documentation tool. The importable module
+# stays `rtldoc` (see [tool.setuptools] below) and the installed CLI
+# command stays `rtldoc` (see [project.scripts]) -- neither has to match
+# the PyPI distribution name.
+name = "pdf-rtldoc"
+version = "1.0.6"
+description = "Geometry-first PDF parser for RTL and complex layouts (Arabic textbooks, financial tables, multi-column docs)"
+readme = "README.md"
+license = { text = "MIT" }
+requires-python = ">=3.10"
+dependencies = [
+    "pymupdf>=1.24",
+    "numpy",
+]
+
+[project.optional-dependencies]
+# cvfallback.py (pixel-based fallback for scanned/flattened pages) isn't
+# wired into the pipeline yet -- nothing in rtldoc/ imports cv2 today.
+cv = ["opencv-python-headless"]
+# ocr.py's scanned-page fallback shells out to the `tesseract` CLI binary
+# directly (apt/brew install tesseract-ocr) rather than the pytesseract
+# Python wrapper -- no pip package needed, so there's no extras group for
+# it. Optional at runtime: if the binary isn't on PATH, a scanned page just
+# gets no blocks, same as before this feature existed.
+# Only tests/make_fixture.py (the synthetic-PDF generator for local testing)
+# needs these; the parser itself never shapes or reverses text on its own.
+dev = ["arabic-reshaper", "python-bidi"]
+
+[project.urls]
+Homepage = "https://github.com/umerjavaidkh/rtldoc"
+Repository = "https://github.com/umerjavaidkh/rtldoc"
+Issues = "https://github.com/umerjavaidkh/rtldoc/issues"
+
+[project.scripts]
+rtldoc = "rtldoc.cli:main"
+
+[tool.setuptools]
+packages = ["rtldoc"]

pdf_rtldoc-1.0.6/rtldoc/__init__.py

@@ -0,0 +1 @@
+"""rtldoc — a geometry-first parser for RTL complex-layout PDFs."""