What Is Docling

The problem docling solves

Feeding a document to an LLM usually means turning it into text first. For a .txt file that's trivial. For a real PDF report with a table, a chart, and a two-column layout, it is not: the raw text extracted by a naive PDF reader often comes out with headings mixed into body text, table cells flattened into one unreadable line, and reading order scrambled by columns. docling exists to solve that specific problem: it runs a trained layout model over the page image, understands where headings, paragraphs, tables, and pictures actually are, and produces one clean structured document from that understanding.

Where docling sits next to markitdown and liteparse

All three libraries answer "turn this document into something an LLM can read", but they sit at different points on a fidelity-versus-speed axis:

• markitdown optimizes for breadth and simplicity: the widest format coverage, a single .convert() call, minimal setup.
• liteparse optimizes for PDF fidelity at speed: layout, forms, and OCR, tuned to stay fast.
• docling goes deepest on structure: a real layout model, a dedicated table-structure model (TableFormer), OCR, and formula and code recognition, at the cost of heavier dependencies and slower conversions.

None of that makes docling strictly "better", it makes it the right tool when structural fidelity (a table your code can actually read as a table, not a wall of text) matters more than raw speed. Lesson 17 tests this claim directly once you've used all three enough to judge for yourself.

One converter, many formats

The core object is DocumentConverter. It needs no arguments to construct, and no API key or network account, everything it does runs locally:

from docling.document_converter import DocumentConverter

converter = DocumentConverter()

InputFormat is the enum listing every format docling's backends know how to read, PDF, DOCX, PPTX, HTML, images, and quite a few more. This course mostly uses pdf, docx, and pptx, but the same DocumentConverter handles all of them without any per-format setup from you.