Skip to main content
docling_architecture

​
Overview

Docling’s architecture is built on a modular design that separates concerns between document parsing, processing pipelines, and output generation. The system is designed to handle multiple document formats through a unified interface while maintaining flexibility and extensibility.

​
Core Components

The architecture consists of four main components that work together to convert documents:

​
1. Document Converter

The DocumentConverter class (docling/document_converter.py:198) is the main entry point for all document conversions. It:
  • Manages format-specific configurations through FormatOption mappings
  • Routes documents to appropriate backends and pipelines based on input format
  • Caches initialized pipelines for performance (keyed by pipeline class and options hash)
  • Handles both single-document (convert()) and batch conversion (convert_all())
from docling.document_converter import DocumentConverter

converter = DocumentConverter()
result = converter.convert("document.pdf")

​
2. Backends

Backends are format-specific parsers that extract raw content from input documents. See Backend Architecture for details. Key backend types:
  • Declarative backends: Directly produce DoclingDocument (DOCX, HTML, Markdown, etc.)
  • Paginated backends: Provide page-level access for pipeline processing (PDF, images)

​
3. Pipelines

Pipelines orchestrate the conversion process, applying ML models and transformations to create structured output. See Pipeline Concepts for details. Main pipeline types:
  • SimplePipeline: For declarative backends that output documents directly
  • StandardPdfPipeline: Multi-threaded PDF processing with OCR, layout analysis, and table extraction
  • VlmPipeline: Vision-language model based conversion
  • AsrPipeline: Audio transcription and conversion

​
4. DoclingDocument

The unified document representation format that all conversions produce. See DoclingDocument for details.

​
Conversion Flow

The typical document conversion follows this flow:
1

Format Detection

The DocumentConverter identifies the input format and retrieves the corresponding FormatOption configuration.
2

Backend Initialization

A format-specific backend is instantiated to parse the input document (e.g., DoclingParseDocumentBackend for PDFs).
3

Pipeline Execution

The pipeline orchestrates the conversion:
  • Build: Extract content using the backend
  • Assemble: Structure content into a DoclingDocument
  • Enrich: Apply ML models for enhancement (optional)
4

Output Generation

The ConversionResult contains the final DoclingDocument and metadata, ready for export or further processing.

​
Pipeline Stages

For complex formats like PDF, pipelines execute multiple stages:
  1. Page Initialization: Load page-level backends and extract raw content
  2. Preprocessing: Scale images, prepare for model input
  3. OCR: Text recognition for scanned or image-based content
  4. Layout Analysis: Detect document structure (headings, paragraphs, tables, figures)
  5. Table Structure: Parse table cells and relationships
  6. Assembly: Combine page elements into a unified document
  7. Enrichment: Apply optional models (picture classification, chart extraction, etc.)
The StandardPdfPipeline executes these stages in parallel across multiple threads for optimal performance. Pages are processed in batches with configurable concurrency.

​
Format-to-Pipeline Mapping

Each input format has a default backend and pipeline configuration:
FormatBackendPipelinePurpose
PDFDoclingParseDocumentBackendStandardPdfPipelineFull PDF processing with OCR and layout
IMAGEImageDocumentBackendStandardPdfPipelineImage-based document conversion
DOCXMsWordDocumentBackendSimplePipelineMicrosoft Word documents
HTMLHTMLDocumentBackendSimplePipelineWeb content
MDMarkdownDocumentBackendSimplePipelineMarkdown files
AUDIONoOpBackendAsrPipelineSpeech-to-text transcription
You can override the default mappings by providing custom format_options when initializing DocumentConverter. This allows you to use different backends or pipeline configurations for specific formats.

​
Extensibility

Docling’s architecture supports extension through:
  • Custom backends: Implement AbstractDocumentBackend for new formats
  • Custom pipelines: Subclass BasePipeline for specialized processing
  • Plugin models: Add ML models via the plugin system
  • Custom serializers: Define new export formats
See the Plugins documentation for details on extending Docling.

​
Performance Considerations

​
Pipeline Caching

The DocumentConverter caches initialized pipelines using a composite key of (pipeline_class, options_hash). This means:
  • Pipelines with identical configurations are reused across documents
  • Heavy ML models are loaded once per pipeline instance
  • Thread-safe access ensures concurrent conversions can share pipelines

​
Batch Processing

For optimal throughput when processing multiple documents: 
converter = DocumentConverter()
sources = ["doc1.pdf", "doc2.pdf", "doc3.pdf"]

for result in converter.convert_all(sources):
    # Process results as they complete
    print(f"Converted: {result.input.file}")
Batch processing enables:
  • Document-level parallelism (configurable via settings.perf.doc_batch_concurrency)
  • Efficient resource utilization
  • Pipeline sharing across documents

​
References

Docling Technical Report

Deep dive into Docling’s architecture and design decisions

Backend Architecture

Learn about format-specific document backends

Pipeline Concepts

Understand pipeline orchestration and processing

DoclingDocument

Explore the unified document representation

Build docs developers (and LLMs) love

Get started for freeTalk to us