pdfmark-ai 0.4.0

Convert PDF files to high-quality Markdown using LLM vision models

pdfmark-ai

PDF to Markdown, powered by LLM vision.

Drop a PDF, get clean Markdown — tables, formulas, code, figures, all handled.

Demo · Installation · Quick Start · Configuration · CLI Reference

English | 简体中文

pdfmark-ai doesn't parse PDFs the traditional way. Instead, it renders each page as an image and lets multimodal LLMs (Claude, Kimi, Qwen, etc.) "read" it — just like a human would. The result? Clean, structured Markdown that handles what other tools simply can't: complex tables with merged cells, inline math formulas, source code blocks, embedded diagrams, and even blurry scans.

Demo

Real conversion results on academic papers and technical documents — no post-editing.

Image Extraction, Tables & Code

PDF original — mixed figures, tables & code	Converted Markdown — images extracted, tables formatted

PDF original — tables & code blocks	Converted Markdown — syntax-highlighted code

PDF original — charts & formulas	Converted Markdown — chart images referenced

Math Formulas & Blurred Content

PDF original — dense math formulas	Converted Markdown — LaTeX `$...$` and `$$...$$` wrapping

PDF original — blurred / low-quality scan	Converted Markdown — content correctly recognized

Features

• 🖼️ Vision-based extraction — treats each page as an image, handles complex layouts that traditional parsers miss
• 🧮 Math formulas — LaTeX rendering with automatic $...$ and $$...$$ wrapping
• 📊 Complex tables — merged cells, multi-row headers, nested structures
• 💻 Code blocks — syntax-appropriate formatting for source code
• ✂️ Image extraction — --crop-images to crop figures and diagrams as separate files
• 🔍 Blur tolerance — handles low-quality and blurred scans with high recognition accuracy
• 🤖 Multi-provider — Claude, Kimi, Xiaomi, Qwen, and any OpenAI-compatible API
• ⚡ Incremental caching — SHA-256 progressive cache avoids re-processing unchanged pages

Installation

pip install pdfmark-ai

Requirements

• Python >= 3.10
• An LLM API key (Anthropic, Kimi, Qwen, or OpenAI-compatible)

Quick Start

# Step 1: Generate config templates in your current directory
pdfmark --init

# Step 2: Edit .env — uncomment ONE provider and fill in your API key
#   e.g. LLM_API_KEY=your-kimi-api-key

# Step 3: Run
pdfmark input.pdf -o output.md

Default provider is Kimi (kimi-for-coding). To use a different provider, either edit .env to set LLM_MODEL / LLM_BASE_URL, or change active_provider in pdfmark.toml.

💡 Tip: Configuration files (.env and pdfmark.toml) are always read from your current working directory — not from the package installation directory. Place them alongside your PDF files or in your project root.

Configuration

pdfmark-ai uses a 4-layer priority chain: CLI args > env vars > TOML config > defaults.

Config files live in your working directory (where you run pdfmark):

File	Purpose	Contains
.env	API keys & overrides	LLM_API_KEY, LLM_MODEL, LLM_BASE_URL
pdfmark.toml	Provider presets & settings	providers, DPI, chunking, caching

You can generate both files with pdfmark --init, or create them manually.

.env (API keys)

# Uncomment ONE provider and add your key:
LLM_API_KEY=your-kimi-api-key
# LLM_API_KEY=your-anthropic-api-key
# LLM_API_KEY=your-qwen-api-key

# Optional: override model or base URL
# LLM_MODEL=claude-sonnet-4-20250514
# LLM_BASE_URL=https://api.your-provider.com/v1

pdfmark.toml (settings)

active_provider = "anthropic"

[providers.anthropic]
base_url = "https://api.anthropic.com"
model = "claude-sonnet-4-20250514"

[render]
dpi = 150

[cache]
enabled = true
dir = "~/.cache/pdfmark"

Supported Providers

Provider	active_provider	Notes
Anthropic Claude	anthropic	Supports Opus 4.6, Sonnet 4.6 and other Claude models. Uses Anthropic Messages API natively.
Kimi (Moonshot)	kimi	Anthropic-compatible API
Xiaomi (MiMo)	xiaomi	Auth token required
Qwen (Alibaba)	qwen	OpenAI-compatible SDK
Any OpenAI-compatible	set LLM_BASE_URL	Set LLM_SDK_TYPE=openai

CLI Reference

Usage: pdfmark [OPTIONS] [INPUT]

Arguments:
  INPUT                   Path to the PDF file to convert

Options:
  --init                  Generate .env and pdfmark.toml config templates
  -f, --force             Overwrite existing config files (use with --init)
  -o, --output            Output markdown file path
  --lang                  Document language (e.g. 'en', 'zh', 'auto')
  --crop-images           Extract visual regions from pages as images
  --refine                Run optional LLM global refinement pass
  --no-cache              Disable caching of rendered pages and chunks
  --no-frontmatter        Omit YAML frontmatter from output
  --detect-only           Detect document structure and print sections
  --config                Path to a TOML configuration file
  --dpi                   Rendering DPI for PDF pages
  --model                 LLM model identifier
  --api-key               LLM API key (or set LLM_API_KEY env var)
  --base-url              LLM API base URL
  --max-concurrent        Maximum concurrent LLM requests

Image Extraction

Use --crop-images to extract figures and diagrams from the PDF as separate image files:

pdfmark input.pdf -o output.md --crop-images

Cropped images are saved alongside the output file (e.g., images/page_003_fig_001.png). Crop mode and plain mode use separate caches, so you can switch freely without needing --no-cache.

License

MIT