Docling

Docling is integrated into the Data Prep Kit open-source toolkit to facilitate the preparation of unstructured data for LLM applications. It is used for large-scale data processing, ranging from local laptop environments to datacenter scales.

Key integration workflows include:

• PDF ingestion to Parquet: Converting PDF documents into Parquet format for efficient storage and processing.
• Document chunking: Breaking down documents into smaller, manageable chunks suitable for LLM contexts.

Docling supports multiple Optical Character Recognition (OCR) engines. These are not included in the core package and must be installed as extra packages depending on your requirements:

• RapidOCR
• Nemotron-OCR
• EasyOCR
• ocrmac
• tesseract-CLI
• tesserocr

Docling allows you to enrich the conversion pipeline with additional steps that process specific document components like code blocks, formulas, or pictures. These enrichments are disabled by default because they require additional model executions which increase processing time.

The Docling usage skill is structured to allow AI agents to load only the specific documentation required for a task. The entry point is SKILL.md, which routes the agent to specialized reference files:

| Task | Reference File | |---|---|---| | Read/convert a file from the shell | cli.md | | Convert in code and tune the pipeline (PipelineOptions) | python-sdk.md | | Extract specific typed fields from a document | extraction.md | | Chunk documents for RAG | rag.md | | Offload conversion to a remote service | service-client.md | | Install only the dependencies you need | slim-packaging.md |

These files are located within the installed package at docling/.agents/skills/docling/references/.

Introduced in Docling v2, DoclingDocument is a unified document representation format defined as a Pydantic datatype. It provides a structured way to express document features including:

• Content Types: Text, Tables, Pictures, and more.
• Hierarchy: Document structure using sections and groups.
• Disambiguation: Separation between the main document body and furniture (headers, footers, etc.).
• Layout: Bounding box information for items (if available).
• Provenance: Information regarding the origin of the content.

The Pydantic type definitions are located in the docling_core.types.doc module.

Docling provides different pipeline implementations for document processing:

• StandardPdfPipeline: A comprehensive pipeline designed for high-fidelity PDF parsing and structural analysis.
• SimplePipeline: A lighter-weight pipeline option for faster or less complex conversion tasks.

When using picture-description enrichment, you can select models using picture_description_preset (for named presets) or picture_description_custom_config (for full control).

Note: The parameters picture_description_local and picture_description_api are deprecated as of docling-serve v1.21.0.

To use remote OpenAI-compatible API endpoints for description, the server must be launched with DOCLING_SERVE_ENABLE_REMOTE_SERVICES=true.

When using the Tectonic engine for TikZ rendering, Docling handles failures gracefully:

1. Success: If Tectonic compilation succeeds, the TikZ diagram is rasterized and stored as an image.
2. Failure: If compilation fails, times out, produces no PDF, or rasterization fails, Docling does not drop the figure. Instead, it preserves the original TikZ source as fallback code metadata.

Docling provides three interchangeable Automatic Speech Recognition (ASR) backends. You can choose between them based on your hardware and performance requirements:

1. Native Whisper (openai-whisper): The default backend. Works on CPU and CUDA. Offers the broadest compatibility.
2. MLX Whisper (mlx-whisper): Optimized specifically for Apple Silicon (M-series Macs) using MPS.
3. WhisperS2T (whisper-s2t-reborn): An experimental, high-throughput backend using CTranslate2. It uses batched, VAD-segmented decoding and is typically the fastest on CPU and CUDA, using less VRAM for large models. Note: Not available on Apple Silicon.

To use these, ensure you have installed the asr extra during installation.

A chunker is a Docling abstraction that operates directly on a DoclingDocument. It returns a stream of chunks, where each chunk contains a string representation of a document segment and its associated metadata.

There are two conceptual ways to chunk documents in a Docling workflow:

1. Post-processing: Exporting a DoclingDocument to Markdown and then using external tools to chunk the text.
2. Native Chunking: Using Docling's built-in chunker classes to operate directly on the document structure. This is the approach covered by the native chunker API.

Docling's chunkers follow a common interface (BaseChunker), making them compatible with GenAI frameworks like LlamaIndex. You can use built-in chunkers, your own custom implementations, or third-party ones as long as they implement the BaseChunker interface.

Docling uses a hierarchy of format options to customize how different file types are processed. You can provide specific configuration objects to the DocumentConverter to tune the extraction behavior for PDF, Word, PowerPoint, Markdown, AsciiDoc, or HTML files.

from docling.document_converter import DocumentConverter, PdfFormatOption

# Example: Customizing PDF conversion options
converter = DocumentConverter(
    format_options={
        "pdf": PdfFormatOption(
            # configuration parameters here
        )
    }
)