Docling

Accepts PDFs, DOCX, PPTX, images, and HTML as input and routes each format through specialized parsers that normalize to an intermediate representation before final structured output. Uses format-specific libraries (PyPDF2/pdfplumber for PDFs, python-docx for DOCX, etc.) with a common abstraction layer that ensures consistent downstream processing regardless of source format.

Applies OCR to scanned documents and images using Tesseract or cloud-based vision APIs, with spatial awareness of text bounding boxes and reading order. Reconstructs logical text flow from detected character positions rather than naive top-to-bottom extraction, preserving document structure and column layouts during text recovery.

Provides confidence scores and quality metrics for extracted elements, particularly from OCR and vision-based extraction. Includes per-element confidence scores (character-level for OCR, element-level for tables/layout) and aggregate metrics to enable downstream systems to assess extraction quality and implement confidence-based filtering or post-processing.

Allows definition of custom element types and processing logic through a plugin or extension mechanism, enabling teams to extend Docling for domain-specific document types (e.g., medical forms, financial statements) without modifying core code. Supports custom extraction rules, validation, and export formats tailored to specific use cases.

Splits extracted document structure into chunks suitable for RAG systems, respecting semantic boundaries (paragraphs, sections, tables) rather than naive character-count splitting. Implements configurable chunk size, overlap, and boundary detection to preserve semantic coherence while enabling efficient retrieval. Maintains chunk metadata (source page, section, confidence) for traceability.

Identifies table regions within documents using computer vision or heuristic-based detection, then parses table structure (rows, columns, merged cells) and extracts cell content with semantic understanding. Outputs tables as structured data (JSON, CSV, or pandas DataFrames) with metadata about cell types, headers, and relationships.

Analyzes the spatial arrangement of document elements (text blocks, images, tables, headers, footers) and reconstructs logical document structure including reading order, hierarchy, and semantic roles. Uses computer vision techniques (connected component analysis, bounding box clustering) combined with heuristics to identify sections, subsections, and element relationships.

Converts extracted document structure to Markdown format with preservation of heading hierarchies, emphasis (bold/italic), lists, code blocks, and table formatting. Maps document semantic roles (heading levels, emphasis, list types) to corresponding Markdown syntax, enabling round-trip compatibility with Markdown-aware tools.

Exports parsed documents to JSON format with complete metadata including element types, bounding boxes, confidence scores, and semantic roles. Preserves hierarchical structure and relationships between elements (e.g., table headers, list nesting) in a machine-readable format suitable for downstream processing and integration with other tools.

Defines an internal structured representation (DoclingDocument) that models documents as hierarchical trees of typed elements (TextBlock, Table, Image, etc.) with metadata, relationships, and spatial information. Serves as the canonical intermediate format that all exporters (Markdown, JSON) consume, enabling consistent processing regardless of input format or output target.

Supports processing multiple documents in sequence or parallel with configurable pipeline stages (parsing, OCR, table extraction, layout analysis, export). Allows selective enabling/disabling of stages and custom stage ordering to optimize for specific use cases (e.g., skip OCR for native PDFs, prioritize speed over accuracy).

Identifies and extracts images embedded in documents (PDFs, DOCX, PPTX) while preserving spatial metadata including position, size, and context within the document. Outputs images as separate files with references in the document structure, enabling downstream systems to access both image content and its relationship to surrounding text.

For DOCX and PPTX files, parses the underlying XML structure directly rather than relying on OCR or vision-based extraction. Extracts text, formatting, and structure from Office Open XML format while preserving semantic information (styles, lists, tables) that is encoded in the XML, avoiding information loss from format conversion.