docberry 0.1.0

Extract structured Markdown, tables, figures, and equations from scientific PDFs with proper reading order.

DocBerry

Extract structured Markdown, tables, figures, and equations from scientific PDFs with proper reading order.

DocBerry wraps Docling with a reading-order segmentation layer for two-column academic papers, multi-format asset extraction (PNG/PDF/SVG), and pluggable equation LaTeX enrichment backends.

---

Features

• Reading-order segmentation — Detects full-width headers, two-column body text, and full-width figures/tables, then reorders them into natural reading order.
• Structured Markdown output — Tables replaced with image hyperlinks, equations with LaTeX blocks and image references.
• Asset extraction — Tables (PNG + CSV), figures (PNG + PDF + SVG), equations (PNG + PDF + SVG + LaTeX .txt).
• Equation enrichment — Choose from 4 backends: none, pix2tex (fast), Qwen3.5-0.8B VLM, or Docling CodeFormulaV2.
• Auto-segment + convert — One-shot pipeline that segments then converts.

---

Installation

Requires Python >= 3.10.

# Core (segmentation + conversion + asset extraction)
pip install docberry

# With lightweight equation LaTeX OCR (pix2tex)
pip install 'docberry[pix2tex]'

# With Qwen3.5-0.8B VLM equation enrichment
pip install 'docberry[qwen]'

# With debug overlays (opencv)
pip install 'docberry[debug]'

# Everything
pip install 'docberry[all]'

Pre-download model weights

Models auto-download on first use. To pre-download:

# Core Docling models only
docberry download-models

# All models (Docling + pix2tex + Qwen)
docberry download-models --all

# Specific extras
docberry download-models --pix2tex
docberry download-models --qwen

---

Quick Start

Python API

from docberry import segment_pdf, convert_document

# Step 1 (optional): segment a two-column PDF into reading order
segment_pdf("paper.pdf", "paper_segmented.pdf")

# Step 2: convert to Markdown with asset extraction
result = convert_document(
    "paper_segmented.pdf",
    output_dir="output/",
    extract_assets=True,
    equation_enrichment="pix2tex",
)
print(result.markdown_path)   # output/paper_segmented.md
print(result.tables)          # number of tables extracted
print(result.figures)         # number of figures extracted
print(result.equations)       # number of equations extracted
print(result.elapsed_seconds) # total processing time

# One-shot: segment + convert in a single call
result = convert_document(
    "paper.pdf",
    output_dir="output/",
    extract_assets=True,
    auto_segment=True,
    equation_enrichment="pix2tex",
)

CLI

# Full pipeline (segment + convert)
docberry convert paper.pdf -o output/ --extract-assets --auto-segment

# With equation enrichment
docberry convert paper.pdf -o output/ --extract-assets --equation-enrichment pix2tex

# Convert only (no segmentation)
docberry convert paper.pdf -o output/ --extract-assets

# Segment only
docberry segment paper.pdf -o paper_segmented.pdf

# Segment with debug overlays
docberry segment paper.pdf -o paper_segmented.pdf --debug-dir overlays/

---

CLI Reference

docberry convert

Convert a document to Markdown/JSON with optional asset extraction.

Flag	Default	Description
source (positional)	—	Path to a local file or URL
--output-dir, -o	<stem>_output/	Output directory
--format	markdown	Output format: markdown or json
--extract-assets	off	Extract tables, figures, equations
--layout-model	heron	heron, egret-medium, egret-large, egret-xlarge
--pipeline	standard	standard or vlm
--equation-enrichment	none	none, pix2tex, qwen, docling
--auto-segment	off	Segment PDF for reading order first

docberry segment

Segment a two-column PDF into reading-order pages.

Flag	Default	Description
input (positional)	—	Input PDF path
--output, -o	segmented_output.pdf	Output segmented PDF path
--pages	all	Page spec, 0-based (e.g. 0,2-4)
--debug-dir	—	Directory for debug overlay images

Layout tuning flags: --line-merge-gap, --band-merge-gap, --block-padding, --segment-padding, --min-text-height, --line-split-gap-x, --num-bins, --single-coverage-threshold, --min-side-coverage, --min-center-gap-ratio, --min-band-height-ratio, --caption-merge-gap, --hrule-coverage.

docberry download-models

Pre-download model weights.

Flag	Description
--all	Download all models (Docling + pix2tex + Qwen)
--pix2tex	Download pix2tex weights only
--qwen	Download Qwen3.5-0.8B weights only

---

Python API Reference

segment_pdf(input_pdf, output_pdf, page_spec, debug_dir, config)

Segment a PDF into reading-order pages.

Parameter	Type	Default	Description
input_pdf	str	—	Path to source PDF
output_pdf	str	"segmented_output.pdf"	Output path
page_spec	str | None	None	Page range, e.g. "0,2-4"
debug_dir	str | None	None	Debug overlay directory
config	LayoutConfig | None	None	Tuning parameters

Returns: list[Segment]

convert_document(source, output_dir, ...)

Convert a document to Markdown/JSON with optional asset extraction.

Parameter	Type	Default	Description
source	str	—	File path or URL
output_dir	str | None	auto	Output directory
output_format	str	"markdown"	"markdown" or "json"
extract_assets	bool	True	Extract tables/figures/equations
layout_model	str	"heron"	Layout model name
pipeline	str	"standard"	"standard" or "vlm"
equation_enrichment	str	"none"	"none", "pix2tex", "qwen", "docling"
auto_segment	bool	False	Run segmentation first

Returns: ConversionResult with .markdown_path, .tables, .figures, .equations, .elapsed_seconds

---

Equation Enrichment Comparison

Method	Speed	Quality	Model Size	Extra Deps
none	—	No LaTeX (images only)	—	—
pix2tex	Fast (~1s/eq)	Good for simple equations	~100 MB	pix2tex
qwen	Medium (~3s/eq)	Good, handles complex notation	~1.6 GB	torch, transformers
docling	Slow (~5s/eq)	High (CodeFormulaV2 VLM)	~2 GB	Built into Docling

---

Output Structure

output/
  paper.md                    # Structured Markdown
  tables/
    table-1.png / .pdf / .svg
    table-1.csv
    table-1_caption.txt
  figures/
    figure-1.png / .pdf / .svg
    figure-1_caption.txt
  equations/
    equation-1.png / .pdf / .svg
    equation-1_latex.txt

---

Acknowledgments

• Docling — Core document conversion engine
• pix2tex — Lightweight LaTeX OCR
• Qwen — Vision-language model for equation extraction
• PyMuPDF — PDF processing for layout segmentation

---

License

MIT — see LICENSE.

---

Website | GitHub