Docling

Docling - Universal Document Converter

[!IMPORTANT] Docling is the specialized tool for the "Data Curation" step of RAG (Retrieval Augmented Generation). It turns complex, messy file formats into clean, structured Markdown that LLMs love.

1. Core Capabilities

• Universal Parsing: Handles PDF, DOCX, PPTX, Images, HTML, AsciiDoc, Markdown, and Audio (via Whisper).
• Layout Awareness: Understands tables, headers, footers, page numbers, and reading order in PDFs (OCR included).
• Hybrid Chunking: Smart chunking strategy that uses embedding models to keep semantically related text together, rather than just splitting by character count.
• Export to Markdown: The gold standard format for LLM context.

2. Quick Start

Installation

pip install docling
# Optional: Audio support
pip install ffmpeg-python openai-whisper

Basic Extraction (PDF to Markdown)

from docling.document_converter import DocumentConverter

source = "path/to/document.pdf"  # Can be URL or local path
converter = DocumentConverter()
result = converter.convert(source)

# Export to Markdown
markdown_output = result.document.export_to_markdown()
print(markdown_output)

3. Advanced Usage: Hybrid Chunking

This is the killer feature for RAG. Don't just split by characters; split by meaning.

from docling.document_converter import DocumentConverter
from docling.chunking import HybridChunker

doc_converter = DocumentConverter()
conv_res = doc_converter.convert("complex_report.pdf")
doc = conv_res.document

# Initialize Hybrid Chunker
chunker = HybridChunker(
    tokenizer="intfloat/multilingual-e5-base" # Or other embedding model tokenizer
)

# Generate Chunks
chunk_iter = chunker.chunk(doc)
chunks = list(chunk_iter)

for chunk in chunks:
    print(f"Chunk Text:\n{chunk.text}\n---\n")
    # Store 'chunk.vector' in your DB if using embedded creation here

4. Working with Audio (Transcription)

Docling streamlines the audio-to-text pipeline for RAG.

# Requires: pip install openai-whisper
from docling.document_converter import DocumentConverter

converter = DocumentConverter()
result = converter.convert("meeting_recording.mp3")

# Output includes timestamps and speaker separation if supported
print(result.document.export_to_markdown())

5. RAG Integration Pattern

1. Ingest: Use DocumentConverter to read files from your source directory.
2. Process: Convert every file (PDF, Word, Audio) into a uniform DoclingDocument.
3. Chunk: Apply HybridChunker to split documents intelligently.
4. Embed & Store: Send chunks to your Vector DB (PGVector, Pinecone, etc.).
5. Retrieve: Query your DB for relevant chunks and feed them to the LLM.

Resources

• Docling GitHub & Docs
• Recommended for robust Parsing vs. "Crawl4AI" for Web Parsing.