Marker

Converts PDF, PowerPoint, Word, Excel, EPUB, and image files into a unified internal document representation through a pluggable provider architecture. Each provider handles format-specific extraction (e.g., PDF uses pdfplumber or PyPDF2, Office formats use python-pptx/python-docx), normalizing diverse input types into a common block-based schema for downstream processing. The provider pattern enables extensibility without modifying core pipeline logic.

Uses pre-trained deep learning models (via detectron2 or similar vision transformers) to identify document structure elements (text regions, tables, figures, headers, footers) and their spatial relationships through polygon-based bounding box detection. The layout builder constructs a hierarchical block tree that preserves 2D positioning information, enabling accurate reconstruction of document structure even in complex multi-column or non-linear layouts. This approach outperforms rule-based heuristics for varied document designs.

Processes multiple documents in parallel using a configurable batch pipeline that distributes work across available GPUs or CPU cores. Implements job queuing, progress tracking, and error handling for large-scale document conversion. Supports distributed processing via Python multiprocessing or async I/O, with configurable batch sizes and worker counts. Enables efficient processing of document collections for RAG systems or data extraction pipelines.

Provides a centralized configuration system that manages model selection, processing options, LLM provider credentials, and output format settings. Supports environment variable overrides for deployment flexibility, YAML/JSON configuration files for complex setups, and dynamic component discovery via entry points. Enables users to customize behavior (e.g., which layout model to use, OCR provider, LLM service) without code changes.

Provides a REST API server (FastAPI-based) that exposes document conversion as HTTP endpoints, enabling integration with external systems and web applications. Supports file upload, conversion with configurable options, and streaming output. Implements request queuing, timeout handling, and resource limits to prevent abuse. Enables Marker to be deployed as a microservice for document processing pipelines.

Detects form fields (text inputs, checkboxes, radio buttons, dropdowns) using layout analysis and specialized form processors. Extracts field values and metadata (field name, type, position, default value) and outputs structured data (JSON, CSV) suitable for downstream processing. Supports both filled and unfilled forms, with optional LLM-based field value correction for low-confidence extractions.

Performs optical character recognition (OCR) on document regions where native text extraction fails, using Tesseract or cloud-based OCR APIs as fallback. Integrates text line detection models to identify individual text lines and their bounding boxes, enabling character-level positioning for accurate reconstruction. The system automatically routes content through OCR when PDF text extraction yields low confidence or when processing scanned/image-based documents, with configurable confidence thresholds.

Detects table regions via layout analysis, extracts cell content through OCR or native text extraction, and reconstructs table structure (rows, columns, merged cells) using heuristic-based cell alignment and optional LLM-based refinement. The table processor handles complex tables with merged cells, nested headers, and irregular layouts by analyzing cell boundaries and content relationships. LLM processors can be invoked to correct misaligned cells or infer missing content, trading latency for accuracy.

Detects mathematical expressions (both inline and display equations) using layout analysis and specialized processors that convert LaTeX, MathML, or image-based equations into Markdown-compatible notation (e.g., `$...$` for inline, `$$...$$` for display). Handles both native PDF equations and image-based math through OCR fallback. The system preserves equation positioning and context within document flow.

Detects and extracts images from documents, preserving them as separate files with configurable formats (PNG, JPG, WebP) and resolution. Tracks image metadata (position, size, caption, alt-text) and maintains references in output Markdown/JSON, enabling downstream processing or LLM-based image description. Supports batch image extraction with deduplication to avoid storing identical images multiple times.

Identifies and removes repetitive page elements (headers, footers, page numbers, watermarks) using spatial analysis and content matching heuristics. The system detects elements that appear on multiple pages in similar positions, marks them as artifacts, and excludes them from output. Configurable thresholds allow tuning sensitivity to balance between removing true artifacts and preserving legitimate content that happens to repeat.

Represents documents as a tree of nested blocks (pages, paragraphs, text lines, tables, figures) with spatial metadata (polygon coordinates, bounding boxes, rotation). Each block tracks its type, content, and relationships to parent/sibling blocks, enabling efficient querying and processing of specific element types. The schema supports multiple extraction methods per block type and enables spatial indexing for fast region-based lookups.

Optionally invokes Large Language Models (OpenAI, Anthropic, local models) to refine extracted content, correct OCR errors, improve table structure, generate image descriptions, or fix complex formatting. Implements parallel LLM processing to handle multiple blocks concurrently, with configurable batch sizes and rate limiting. Supports specialized LLM processors for different content types (tables, forms, handwriting, complex layouts), enabling targeted accuracy improvements without processing entire documents through LLMs.

Renders processed documents to multiple output formats (Markdown, JSON, HTML) with configurable options for each format. The renderer system is pluggable, allowing custom renderers for domain-specific formats. Markdown output preserves structure through heading levels, lists, and code blocks; JSON output includes full metadata and spatial information; HTML output enables web-based viewing. Each renderer can be configured to include/exclude specific elements (images, tables, equations, metadata).