Docling

Open-source document parsing library from IBM Research Zurich. Converts PDFs and other document formats into structured representations suitable for RAG pipe…

Docling

Open-source document parsing library from IBM Research Zurich. Converts PDFs and other document formats into structured representations suitable for RAG pipelines, LLM ingestion, and agentic workflows.

Core abstraction

All inputs parse to a DoclingDocument — a structured intermediate object from which you export Markdown, JSON, HTML, or DocTags. Programmatic access to individual components (tables, figures, text blocks) is available before export.

Quick start

from docling.document_converter import DocumentConverter

converter = DocumentConverter()
doc = converter.convert("https://arxiv.org/pdf/2408.09869").document
print(doc.export_to_markdown())

# CLI
docling https://arxiv.org/pdf/2206.01062

# VLM pipeline (Apple Silicon MLX accelerated)
docling --pipeline vlm --vlm-model granite_docling path/to/paper.pdf

Key capabilities

• Layout-aware PDF parsing: reading order, multi-column, headers/footers
• Table structure extraction (rows, columns, multi-level headers)
• Formula → LaTeX conversion
• Figure classification and description generation
• OCR for scanned PDFs
• Audio transcription (ASR)
• Local / air-gapped execution

Integration

• Python: pip install docling
• MCP server: docling-mcp (agent-accessible without custom code)
• Framework integrations: LangChain, LlamaIndex, Haystack, Crew AI

vs. Firecrawl

	Docling	Firecrawl
Primary target	Local files + PDFs	Web URLs
Strength	High-fidelity document structure	HTML → clean markdown
OCR	Yes (scanned PDFs)	No
Table extraction	Deep (structure-aware)	Surface-level
Audio/images	Yes	No
Use case	Research papers, reports, contracts	Web pages, crawling, SPA content

Complementary: Firecrawl for web, Docling for files. See Firecrawl.

RAG Pipeline Role

Docling slots at the ingestion layer — before chunking, embedding, or retrieval:

raw documents → Docling → DoclingDocument → Markdown/JSON → chunk → embed → vector store

Chunking guidance: split on ## section headers, not fixed token counts. Keep figure/table blocks intact — splitting mid-table destroys the relational structure. Docling’s block metadata makes section boundaries explicit.

Combined with Contextual Retrieval (prepend context to chunks before embedding), Docling-quality parsing maximizes retrieval precision.

Relation to other tools

• Complements Firecrawl: Firecrawl handles web URLs; Docling handles files and PDFs
• Works upstream of Contextual Retrieval: clean Docling output → better chunk context → better retrieval