Docling
title: docling type: language created: 2026-05-26 last_updated: 2026-05-26 related: ["radar/languages/OpenParse", "radar/tools/Marker", "radar/platforms/Langfuse", "radar/techniques/RagChunkingStrategies"] sources: ["https://github.com/DS4SD/docling"] radar_quadrant: Languages & Frameworks radar_ring: Assess radar_position: inner
IBM's open-source document ingestion library for generative AI pipelines. Converts PDFs, DOCX, XLSX, PPTX, HTML, and images to structured markdown with layout-aware table extraction, heading hierarchy preservation, and figure identification. Designed as the pre-processing layer for RAG pipelines.
Naive PDF extraction (PyPDF2, pdfminer) reads characters in stream order, which destroys the semantic structure of complex documents. A table becomes a flat sequence of cell values; a multi-column layout reads left-to-right across columns rather than down each column; headings lose their hierarchy. The resulting text chunks are incoherent for retrieval and confusing for language models trying to answer questions about them.
Docling applies layout analysis before text extraction. The pipeline:
- Detects structural elements — tables, figures, headings, lists, paragraphs — using a layout model
- Reconstructs reading order for multi-column and complex layouts
- Extracts tables as proper markdown tables (rows and columns intact)
- Identifies figures with bounding boxes; OCR text within figures where applicable
- Outputs a
DoclingDocumentobject with a structured hierarchy, convertible to markdown, JSON, or HTML
Supported input formats: PDF (native and scanned via OCR), DOCX, XLSX, PPTX, HTML, AsciiDoc, and images.
Docling ships as a document loader for LlamaIndex and LangChain. Replacing a naive PDF reader with docling is typically a one-line change in an existing pipeline. IBM uses it internally at scale for enterprise document processing.
Docling complements open-parse (see radar/languages/OpenParse) and Marker (see radar/tools/Marker). Open-parse focuses on PDF chunking with layout preservation; Marker on high-accuracy PDF-to-markdown conversion. Docling's differentiator is broader format support (DOCX, XLSX, PPTX) and direct integration with major RAG frameworks. For table-heavy documents — financial reports, technical specifications, research papers — docling's structured extraction produces materially better chunk quality than text-only approaches.
docling sits in the Assess ring of the Languages & Frameworks quadrant, at inner position. First studied via GitHub (2024-11-03). Document ingestion quality is a first-order determinant of RAG pipeline accuracy; improving it requires no changes to retrieval or generation logic. IBM production use and framework integrations (LlamaIndex, LangChain) lower adoption risk. Inner position reflects a clear trial path (swap in docling as document loader, compare chunk quality on table-heavy PDFs) and direct complementarity with RAG tools already on the radar. Remaining gate before Trial is confirmed chunk quality improvement on a real corpus with table-heavy documents evaluated end-to-end.
