Marker

Extracts content from PDF, PowerPoint, Word, Excel, EPUB, and image files through a pluggable provider architecture that abstracts format-specific extraction logic. Each provider implements a standardized interface to convert source documents into an intermediate representation that feeds into the layout analysis pipeline, enabling consistent processing across heterogeneous document types without format-specific branching in downstream components.

Analyzes document layout using deep learning models to identify spatial relationships between content blocks (text, tables, images, equations) and constructs a hierarchical block-based document schema that preserves 2D positioning via polygon coordinates. The layout builder processes extracted content through layout detection models to segment pages into logical regions, then structures these regions into a tree hierarchy that enables spatial queries and format-aware rendering without losing document geometry information.

Exposes document conversion functionality through a REST API server with endpoints for single-document and batch conversion, status polling, and result retrieval. The API server manages request queuing, handles concurrent conversions with resource limits, and provides streaming responses for large documents or batch operations.

Detects form regions and fields (text inputs, checkboxes, radio buttons, dropdowns) through layout analysis, extracts field labels and values, and optionally uses LLM processors to infer field types and relationships when layout is ambiguous. The form processor outputs structured data (JSON or CSV) mapping field names to extracted values, enabling programmatic access to form data without manual parsing.

Identifies and removes page headers, footers, page numbers, and other document artifacts through layout analysis and heuristic filtering, preserving only main content. The artifact filter uses spatial analysis (e.g., content in top/bottom margins, repeated across pages) and pattern matching to distinguish artifacts from content, improving document quality for downstream processing.

Detects table regions using layout analysis, extracts table content and structure, and optionally uses LLM processors to correct OCR errors, infer missing cell values, and resolve ambiguous table boundaries. The table processor combines computer vision-based table detection with optional LLM-powered post-processing that can handle malformed tables, merged cells, and complex headers by reasoning about table semantics rather than relying solely on grid detection.

Detects mathematical expressions (inline and display equations) within documents using layout analysis, performs OCR on equation regions, and converts recognized formulas to LaTeX notation for accurate Markdown rendering. The system distinguishes between inline math (within text flow) and display equations (block-level), preserving mathematical semantics and enabling proper rendering in Markdown and HTML outputs that support LaTeX.

Performs OCR on text regions and image-based content using configurable OCR engines (Tesseract, EasyOCR, or cloud APIs) with confidence scoring and optional fallback to alternative engines when primary OCR fails. The OCR processor integrates with the layout pipeline to apply OCR only to regions identified as text, preserving spatial context and enabling confidence-based filtering or LLM-powered correction of low-confidence extractions.

Optionally routes document blocks through LLM processors (table correction, form field extraction, handwriting recognition, image description, complex layout handling) to improve conversion accuracy beyond what layout detection and OCR alone can achieve. The system supports parallel LLM processing with configurable batch sizes and provider selection (OpenAI, Anthropic, Ollama, etc.), enabling cost-effective enhancement of specific block types while maintaining performance through batching and async execution.

Renders the hierarchical block structure into multiple output formats (Markdown, JSON, HTML) with format-specific optimizations for each target. The renderer system uses pluggable format handlers that understand format-specific constraints (e.g., Markdown table syntax, HTML semantic tags, JSON schema) and apply appropriate transformations to preserve document semantics while conforming to output format requirements.

Processes multiple documents in parallel using configurable hardware acceleration (CUDA GPUs, CPU, or Apple Metal Performance Shaders) with batch-level optimization and resource management. The batch processor distributes documents across available compute resources, manages model loading and inference scheduling, and provides progress tracking and error handling for large-scale document conversion workflows.

Extracts images from documents, preserves them as separate files with references in output Markdown/HTML, and optionally generates descriptive captions using vision-capable LLMs. The image processor identifies image regions through layout detection, extracts image data, manages file naming and organization, and can enhance accessibility by generating alt-text through LLM vision models.

Provides a centralized configuration system that manages processing parameters (model selection, LLM providers, output formats, hardware preferences) through environment variables, config files, and programmatic APIs. The system uses entry points for component discovery, allowing providers, processors, and renderers to be registered and discovered dynamically without hardcoding dependencies.