Docling

Docling, from IBM Research Zurich and now hosted by the LF AI and Data Foundation, converts documents into structured representations ready for LLM and RAG pipelines. It parses PDF, DOCX, PPTX, XLSX, HTML, EPUB, images, email formats, and even audio and video files, then exports Markdown, HTML, JSON, or its DocTags format while preserving reading order. Its models perform layout analysis, table structure recognition, and formula and code extraction, with OCR for scanned pages and speech recognition for audio inputs; the compact GraniteDocling visual language model at 258M parameters handles end-to-end conversion. Everything runs locally, which matters for sensitive documents, and an API server mode is available. Native integrations cover LangChain, LlamaIndex, Haystack, and CrewAI, plus an MCP server for agent workflows. Requiring Python 3.10 or later and MIT licensed, it is used by developers building document understanding and retrieval pipelines over messy real-world files.