Docling

Docling is an open-source document processing toolkit originally developed by IBM Research that converts documents from virtually any format into clean, structured representations ready for AI consumption. With Apache 2.0 licensing, local execution, and integrations with every major AI framework, it's become one of the most practical tools for teams building RAG systems and document-understanding agents.

Docling handles the formats teams actually encounter: PDF (including scanned), DOCX, PPTX, XLSX, HTML, LaTeX, images (PNG, JPEG, TIFF), and even audio files (WAV, MP3) via automatic speech recognition. Recent releases added WebVTT caption parsing, XBRL financial reports, and USPTO patent documents. This breadth means you don't need separate parsers for each document type — Docling normalizes everything into its unified DoclingDocument format.

PDF parsing is where Docling truly separates from simpler tools like PyPDF or pdfplumber. The Heron layout model (released December 2025) provides faster parsing while accurately detecting page layout, reading order, table structures, code blocks, mathematical formulas, and image classification. It handles multi-column layouts, headers/footers, and complex nested tables that break most other parsers. For OCR on scanned documents, Docling integrates multiple OCR engines and even supports IBM's Granite-Docling-258M vision-language model — a 258M parameter VLM purpose-built for document-to-text conversion that preserves complex layouts in a single inference pass.

Every parsed document converts to the DoclingDocument unified representation, which you can then export as Markdown, HTML, JSON (lossless), WebVTT, or DocTags. The JSON export preserves the full document structure — headings, paragraphs, tables, lists, figures — with coordinates and reading order metadata. This is critical for RAG systems where chunk boundaries and document structure affect retrieval quality. See our guide on building effective RAG systems for why document structure matters.

Docling provides plug-and-play integrations with LangChain, LlamaIndex, CrewAI, and Haystack. These aren't thin wrappers — they're maintained connectors that feed parsed documents directly into each framework's document loaders and chunking pipelines. The MCP server integration (added in 2025) lets any MCP-compatible AI agent use Docling as a document parsing tool, making it accessible from Claude, Cursor, and other MCP clients.

Unlike cloud-based document AI services from Google or Azure, Docling runs entirely locally. Install with pip install docling and process sensitive documents without sending data to any external server. This is essential for healthcare, legal, and financial teams with strict data governance requirements. The CLI makes batch processing straightforward for pipeline automation.

Recent releases added rich metadata extraction capabilities including document language detection, page-level bounding boxes for every element, confidence scores on OCR results, and hierarchical section labeling. The TableFormer model achieves over 90% F1 on complex table structure recognition benchmarks (PubTabNet, FinTabNet), making it among the best open-source options for extracting structured data from tables embedded in PDFs. Docling's chunking utilities — HybridChunker and HierarchicalChunker — leverage this metadata to split documents at semantically meaningful boundaries rather than arbitrary token counts, which measurably improves retrieval precision in RAG systems.

On GPU hardware (e.g., a single NVIDIA A100), Docling processes approximately 10–15 pages per second for standard layout analysis, and 3–5 pages per second when the full VLM pipeline is engaged. CPU-only throughput is roughly 5–10× slower depending on document complexity. The project's GitHub repository has accumulated over 18,000 stars since its public release, reflecting strong community adoption. The SmolDocling model variant (released early 2026) reduced the VLM footprint to under 256M parameters while maintaining competitive accuracy, making GPU requirements more accessible for smaller teams.