docling
docling-project/docling · Python · MIT
Turn messy PDFs and Office docs into clean Markdown, JSON or a layout tree.
この解説はまだ日本語がありません。英語の原文を表示しています。
Docling uses VLM- and rule-based layout analysis to extract text, tables and reading order, ready to feed directly into RAG or fine-tuning pipelines.
スコア
導入の要点
pip install docling 深読み
Docling is IBM's document parser: it turns PDF, DOCX, PPTX, HTML and images into one unified, layout-aware representation — the DoclingDocument — instead of a pile of per-format quirks.
It handles the hard parts of real documents: page layout and reading order, table structure reconstruction, OCR for scanned pages, and figure and code extraction. Clean Markdown or JSON output is why it has become a default front-end for RAG pipelines.
Around the core library sit a CLI, a Python API, an OpenAI-compatible API server and an MCP server; the GraniteDocling 258M vision model covers document understanding, and the DocTags approach behind it is described in an arXiv paper.