Docling

Open-source document parsing for AI: PDFs, Office files, audio and video into clean, structured Markdown/JSON

Free· Free and open source under MIT License. Hosted by the LF AI & Data Foundation; no paid tiers.RAGGraniteDocling 258M and other vision-language models

Visit website →

Best for

ML and platform engineers building RAG systems, doc-QA copilots, or agent pipelines that must ingest heterogeneous documents on their own infrastructure.

Skip if

Non-technical users who want a point-and-click PDF-to-Word converter, or teams unwilling to self-host and manage a Python service.

Docling is an open-source document processing toolkit that turns messy real-world files into clean, structured representations that RAG pipelines and LLM agents can actually reason over. Originally started at IBM Research Zurich and now hosted by the LF AI & Data Foundation, it handles PDF, DOCX, PPTX, XLSX, HTML, EPUB, images, LaTeX, email, and even audio (WAV/MP3) and video, exporting to Markdown, HTML, JSON, WebVTT, DocLang, or its own DocTags format. Its headline capability is advanced PDF understanding: page layout analysis, reading-order recovery, table structure extraction, code and formula recognition, and image classification, driven by purpose-built vision-language models like GraniteDocling (258M). OCR is built in for scanned documents with a choice of engines, and audio is transcribed via ASR. Docling is designed to be dropped straight into modern AI stacks: it ships first-class integrations for LangChain, LlamaIndex, CrewAI, and Haystack, and exposes itself as a Python library, a CLI, an HTTP API server, and an MCP server for use inside agent frameworks. Because everything can run locally, teams handling sensitive, regulated, or air-gapped data can process documents without shipping bytes to a third-party API. Typical workflows include preparing enterprise document corpora for retrieval, converting technical PDFs (research papers, manuals, financial filings in XBRL/JATS) into structured chunks, powering doc-QA copilots, extracting tables from scanned reports, and normalising heterogeneous knowledge bases before embedding. It targets engineers and ML/AI teams building on top of LLMs rather than end users who just want a converter UI.

Editor's take

Docling is the strongest open-source answer we have found to the boring-but-critical problem of turning real PDFs into RAG-ready text. The combination of a small purpose-built VLM, MIT licensing, and drop-in LangChain/LlamaIndex/MCP hooks makes it a default recommendation for anyone building document AI in-house rather than paying per page to a hosted extractor.

— The AI Tool Bible editorial team

Pros

✅ Broad format coverage: PDF, Office, HTML, EPUB, images, LaTeX, email, plus audio and video
✅ Genuinely strong PDF parsing: layout, reading order, tables, formulas, and code blocks preserved
✅ Runs fully local, so sensitive or air-gapped document processing stays on your own infrastructure
✅ Native integrations with LangChain, LlamaIndex, CrewAI, and Haystack shorten the path to a working RAG pipeline
✅ Multiple delivery modes: Python library, CLI, HTTP API server, and MCP server for agents
✅ Purpose-built small VLMs (GraniteDocling 258M) keep GPU/CPU costs modest compared to calling frontier models per page
✅ Permissive MIT license and Linux Foundation governance make it safe for commercial adoption

Cons

⚠️ Python-only public library; non-Python stacks must go through the HTTP or MCP server
⚠️ High-fidelity parsing of complex PDFs benefits from a GPU, which raises the bar for self-hosting
⚠️ No hosted SaaS or managed service: teams must run and maintain their own deployment
⚠️ Documentation and examples assume engineering fluency; there is no non-technical UI
⚠️ Extraction quality on unusual layouts (multi-column scans, handwriting) still varies and may need post-processing

Use cases

RAG document ingestionPDF table extractionScientific paper parsingFinancial filing analysis (XBRL, JATS)Enterprise knowledge base preparationDoc-QA chatbotsAgent-based document workflows via MCPOCR for scanned archivesAudio and video transcription for multimodal RAGConverting Office documents to Markdown for LLMs