/docling

name: docling
description: PRIMARY skill for converting .pdf, .docx, .pptx, .xlsx, .doc, .ppt, .xls, images, and audio/video files (.mp3, .wav, .m4a, .mp4, .mov, etc.) to Markdown. Always prefer this over the docx/pdf/pptx/xlsx/paddleocr skills — docling uses GPU-accelerated OCR + layout detection + table structure extraction for documents, and Whisper ASR for speech-to-text on audio/video. Also supports LaTeX and VTT subtitle files.
version: 1.2.0
metadata:
  requires:
    bins: ["docling"]
    python: ">=3.10"

Docling — Local Document → Markdown Converter

CLI tool that uses AI models (layout detection, OCR, table extraction) to convert documents into clean, structured Markdown.

Supported Formats

**Office**: PDF, DOCX, PPTX, XLSX **Images**: PNG, JPG, TIFF (with OCR) **Web**: HTML, Markdown **Data**: CSV **Academic**: LaTeX, XML (USPTO, JATS, METS) **Media**: Audio/video transcripts (VTT) **Other**: JSON (docling format), ASCIIDoc

When to Use

| Scenario | Tool | |----------|------| | 📄 **PDF / DOCX / PPTX / XLSX on disk** | ✅ **docling** | | 🖼️ Images with text (OCR) | ✅ **docling** | | 📊 CSV / Markdown / HTML on disk | ✅ **docling** | | 🎤 Audio/video transcripts | ✅ **docling** | | 🌐 Public web page → markdown | ❌ Use `WebFetch` (no install needed) | | 🔒 Web page requiring login (WeChat, Zhihu, etc.) | ❌ Use `kimi-webbridge` or Chrome extension | | 🔍 Search for web pages | ❌ Use `WebSearch` |

**Scope boundary**: docling handles **files on disk**. Web content is out of scope — use the right tool for the job.

Prerequisites

> **Molio 用户无需手动装 Python**：Molio 预下载 docling 时，若本机没有 Python ≥3.10，会自动准备一份独立的 Python 运行环境（国内源优先，约 22MB），全程零操作。下面的手动 `pip` 步骤仅作为**回退方案**（例如你在 Molio 之外单独使用 docling，或自动供给因特殊平台/网络失败时）。

Install docling

# China users (recommended) — uses Tsinghua mirror, ~10x faster than default
pip install docling -i https://pypi.tuna.tsinghua.edu.cn/simple

# International users
pip install docling

Verify:

docling --version

⚠️ First-time PDF usage

The first time you convert a PDF, docling downloads ~500MB of AI models (layout + table structure) to `~/.cache/huggingface/`. Subsequent runs reuse cached models and are much faster.

**For users in China**: set the HuggingFace mirror environment variable **before** first run, otherwise model download will timeout:

# Linux / macOS / Git Bash
export HF_ENDPOINT=https://hf-mirror.com
docling file.pdf --to md

# Windows CMD
set HF_ENDPOINT=https://hf-mirror.com
docling file.pdf --to md

# Windows PowerShell
$env:HF_ENDPOINT="https://hf-mirror.com"
docling file.pdf --to md

**Tip**: If you see connection timeout errors while downloading models, this is almost always the cause.

⚠️ First-time audio/video usage

ASR (Whisper) models are **downloaded on demand** — only when you actually convert an audio/video file. Processing PDFs/images does **not** trigger a Whisper download, and vice versa.

• **Location**: same cache as document models, `~/.cache/huggingface/`
• **Size by model**: `whisper_tiny` ~75MB (default), `whisper_small`/`base` ~250-400MB, `whisper_medium` ~1GB, `whisper_large`/`turbo` ~1.5-3GB
• **Reusable**: downloaded once, reused on subsequent runs (no re-download)
• **China users**: the same `HF_ENDPOINT=https://hf-mirror.com` mirror applies — set it before the first audio run, otherwise the download will time out

First run is slow (downloading the model); subsequent runs are pure local inference, speed depends on CPU/GPU.

Quick Commands

Document → Markdown (most common)

# Auto-detect format (PDF, DOCX, PPTX, XLSX, image)
docling "path/to/file.pdf" --to md --output .molio/docling

# Force a specific format when auto-detect fails
docling "file.docx" --from docx --to md --output .molio/docling
docling "slides.pptx" --from pptx --to md --output .molio/docling
docling "report.xlsx" --from xlsx --to md --output .molio/docling

Output: creates a `.md` file in the specified output directory. **统一用 `.molio/docling` 作为输出目录**（Molio 工作区，文件树扫描会跳过，不污染 vault 根）。

PDF with specific options

# PDF with OCR (for scanned documents)
docling "scanned.pdf" --ocr --to md --output .molio/docling

# Use pypdfium2 backend (fallback when default parser fails)
docling "problem.pdf" --pdf-backend pypdfium2 --to md --output .molio/docling

# Force OCR even if text layer exists
docling "mixed.pdf" --force-ocr --to md --output .molio/docling

OCR on images

docling "screenshot.png" --from image --to md --output .molio/docling
docling "photo.jpg" --from image --ocr --to md --output .molio/docling

Audio / Video → Markdown (speech-to-text)

Uses Whisper ASR models to transcribe audio/video files into Markdown.

# Audio file (default whisper_tiny — fast, lower accuracy)
docling "recording.mp3" --from audio --to md --output .molio/docling

# Video file with a more accurate model
docling "meeting.mp4" --from audio --pipeline asr --asr-model whisper_large --to md --output .molio/docling

# Apple Silicon: use MLX variant for faster local inference
docling "interview.m4a" --from audio --pipeline asr --asr-model whisper_medium_mlx --to md --output .molio/docling

# Parse an existing VTT subtitle file (no ASR, text-only)
docling "captions.vtt" --from vtt --to md --output .molio/docling

**Model size guide**:

| Model | Speed | Accuracy | Size | Use case | |-------|-------|----------|------|----------| | `whisper_tiny` (default) | ⚡ fastest | low | ~75MB | quick drafts, English-heavy | | `whisper_small` / `whisper_base` | fast | medium | ~250-400MB | general use | | `whisper_medium` | medium | good | ~1GB | Chinese / noisy audio | | `whisper_large` / `whisper_turbo` | slow | best | ~1.5-3GB | high-accuracy transcripts |

Use `_mlx` suffix on Apple Silicon, `_native` for CPU fallback across platforms.

GPU acceleration

# Use GPU if available (auto-detects CUDA/MPS)
docling "file.pdf" --device auto