Docling

Docling is an open-source document parsing library by IBM Research that converts PDFs, Microsoft Word documents, PowerPoint presentations, Excel spreadsheets, HTML files, and images into structured, LLM-ready formats, preserving the semantic structure (headings, tables, lists, figures) that raw PDF text extraction loses. Standard PDF text extraction strips layout information and produces unstructured text streams where tables become garbled columns and document sections merge without boundaries. Docling reconstructs this structure using deep learning models trained on document understanding, producing output in JSON, Markdown, or HTML formats where the hierarchical document structure is represented as machine-readable elements rather than flat text. The table extraction capability is particularly strong: Docling identifies table boundaries, column headers, and cell relationships in complex multi-column tables across scanned and digital PDFs.

Figure detection identifies images within documents and can extract captions and associated text. Docling runs locally with no external API dependency, models are downloaded and run on the user's hardware. The library integrates natively with LlamaIndex and LangChain for use in RAG pipelines. Available under the MIT license.

Commonly used in enterprise document intelligence systems, RAG applications processing legal or financial documents, and research workflows where the structural fidelity of complex document parsing is critical to downstream answer quality.