structmd — Documents in. Structured Markdown out.

How it works

A strict two-stage pipeline

The VLM never writes your Markdown. It only describes the page as JSON — structmd's deterministic builder turns that JSON into Markdown, so results are reproducible, auditable and editable.

STAGE 1 · EXTRACT

Vision model reads the page

Each page is rendered to an image and sent to any Ollama vision model, which returns a typed, position-aware description of every element.

PyMuPDF render any VLM retry + backoff

ARTIFACT · JSON

Inspectable extraction file

Elements land in a versioned JSON document — headings, paragraphs, tables, lists, images, bounding boxes, reading order. Cache it, diff it, fix it by hand.

content-hash cache human-editable schema-versioned

STAGE 2 · BUILD

Deterministic Markdown builder

Pure-Python rules resolve multi-column layouts, merge cross-page paragraphs, normalize heading levels and emit final Markdown. Same JSON in, same bytes out.

column detection heading remap byte-stable

Why it matters Spot a wrong heading or a missed footnote? Edit the JSON, run structmd doc.pdf --from-json, and get corrected Markdown instantly — without re-running the model.

Features

Built for real pipelines

Everything you need between "I have 400 PDFs" and "they are clean Markdown files".

Multi-column aware

Reading order is resolved geometrically — full-width elements first, then left/right columns interleaved by vertical position.

Fast parallel batch mode

An async worker pool converts entire directories concurrently with live progress, per-file error isolation and resumable caching.

Deterministic by design

Stage 2 is pure Python. Rebuild from cached JSON produces byte-identical output — ideal for CI pipelines and regression tests.

Content-hash caching

Every conversion is keyed by file content hash. Re-run a 500-file batch after fixing one prompt and only changed files hit the model again.

One-command CLI

Convert single files, pick page ranges, rebuild from JSON, or batch whole trees — all from a single structmd command with layered configuration.

Local-first, cloud-capable

Runs against your own Ollama server for full privacy, or against Ollama cloud models when a beefier VLM is convenient. Same command line.

Quick start

From zero to Markdown in three commands

<# install with the extra for your input format>
$ pip install "structmd[pdf]"

<# grab any Ollama vision model>
$ ollama pull qwen3.5-vl

<# convert>
$ structmd thesis.pdf -o thesis.md --pages 1-8

from structmd import StructMDPipeline

with StructMDPipeline(model="qwen3.5-vl") as pipeline:
    doc = pipeline.process("report.pdf", output_path="report.md")
    print(doc.markdown[:200])

# need just the JSON artifact?
extracted = pipeline.extract_only("report.pdf")
extracted.save_json("report.extraction.json")

<# convert every supported file in a tree, 4 workers>
$ structmd batch ./contracts/ -o ./markdown/ --workers 4

<# only page 1 of each file (great for invoices)>
$ structmd batch ./invoices/ -o ./out/ --pages 1

<# 1. extract once>
$ structmd paper.pdf --extract-only -o paper.extraction.json

<# 2. fix anything you dislike in the JSON …>
$ $EDITOR paper.extraction.json

<# 3. rebuild deterministically — the VLM never runs again>
$ structmd paper.pdf --from-json -o paper.md

Compatibility

Formats and models

Input format	Extension(s)	Handled by	Extra needed
PDF documents	.pdf	PyMuPDF page renderer	`pip install "structmd[pdf]"`
Office documents	.docx .pptx .xlsx .odt .ods .odp	LibreOffice headless → PDF	system LibreOffice
Images	.png .jpg .jpeg .webp .tiff .bmp	Straight to the VLM	—

Model family	Examples	Prompt profile
Qwen-VL	qwen2-vl, qwen2.5-vl, qwen3.5-vl	Tuned for Qwen's grounding style
SmolVLM	smolvl2, smolvlm	Compact-model friendly wording
PaliGemma	paligemma, paligemma2	Short-instruction profile
Llama Vision	llama3.2-vision	Llama-native formatting rules
Default	gemma*, minicpm-v, moondream, anything else	Generic robust contract
Ollama cloud	append `:cloud` to any tag	Same profiles, remote inference

Landscape

How structmd compares

	Approach	Editable intermediate	Deterministic rebuild	Local-first	License
structmd	Two-stage: VLM → JSON → rules	Yes — cached JSON	Yes — byte-identical	Yes (Ollama)	MIT
MinerU	Layout detection + OCR models	No	No	Yes	AGPL-3.0
Marker	Deep learning heuristics	No	No	Yes	GPL-3.0 / commercial
py-zerox	VLM per page (OpenAI-centric)	No	No	Cloud API	MIT
LlamaParse	Hosted SaaS API	No	No	Cloud API	Proprietary