Docling — Document Parsing for AI

What it is

Docling is IBM's open-source document parsing library that converts PDFs, Word documents, PowerPoint files, images, and HTML into structured markdown or JSON. It handles complex layouts with tables, headers, figures, and multi-column text, producing clean output suitable for RAG pipelines and LLM ingestion. The library uses AI models for layout understanding and OCR when needed.

Developers building RAG applications, document Q&A systems, or any pipeline that needs to extract structured text from documents benefit from Docling. It replaces fragile PDF parsers with a model-based approach that handles real-world document complexity.

How it saves time or tokens

Docling produces clean, structured output that LLMs can process directly. Without it, developers chain together multiple tools (PDF parsers, table extractors, OCR engines) and write custom post-processing code. Docling handles the entire pipeline in one library call. Clean output also means fewer tokens wasted on HTML tags, layout artifacts, and parsing noise.

How to use

1. Install Docling via pip
2. Create a DocumentConverter instance
3. Convert documents and get structured markdown or JSON

Example

from docling.document_converter import DocumentConverter

converter = DocumentConverter()
result = converter.convert('report.pdf')

# Get markdown output
markdown = result.document.export_to_markdown()
print(markdown)

# Get structured JSON
json_output = result.document.export_to_dict()

# Process multiple files
for path in ['doc1.pdf', 'doc2.docx', 'slides.pptx']:
    result = converter.convert(path)
    print(f'{path}: {len(result.document.pages)} pages')

Related on TokRepo

• AI tools for documents — Browse document processing and parsing tools
• RAG tools — Explore retrieval-augmented generation frameworks

Common pitfalls

• Complex table layouts with merged cells may not parse perfectly; verify output for critical documents
• OCR accuracy depends on image quality; low-resolution scans produce lower quality text extraction
• Processing large documents (100+ pages) can be slow; batch processing with parallel workers improves throughput