Chandra is an open-source (Apache 2.0 / OpenRAIL-M) multimodal visual language model tailored explicitly for advanced document understanding. Instead of outputting unformatted walls of text like traditional optical character recognition (OCR) software, Chandra processes whole document pages at once to analyze layout semantics. It seamlessly converts image scans, forms, and PDFs into highly structured Markdown, HTML, and JSON formats—faithfully reconstructing multi-column flows, complicated tables, and mixed handwriting while isolating graphics alongside their original captions.
Use Cases
High-Fidelity Document Processing for GraphRAG
Convert complex multi-column academic papers, corporate files, and technical textbooks into clean Markdown text blocks to power highly contextual AI retrieval engines.
Structured Financial & Invoice Sheet Extraction
Extract deep data tables from financial disclosures, tax paperwork, or billing invoices, preserving complex structural spans, row relationships, and nested hierarchies perfectly in HTML or JSON.
Mixed Typoscript & Handwriting Digitization
Digitize mixed-medium records—such as annotated legal files, handwritten math equations, or field medical charts—accurately rendering text without failing on written notes.
Automated Physical Form Parsing
Process scanned government, medical, or corporate forms, extracting filled-in blanks and automatically identifying checkbox statuses (checked vs. unchecked) as machine-readable keys.
High-Volume On-Premise Document Pipelines
Deploy scalable batch OCR processes locally using specialized vLLM server clusters, bypassing ongoing remote token bills and preserving sensitive on-prem data privacy boundaries.
Features & Benefits
Full-Page Semantic Region Classification
Identifies and isolates 18+ specific page block types, separating standard text from SectionHeaders, Equations, TableOfContents, Footnotes, Diagrams, and Forms.
Outperforming Commercial AI Models
Scores 83.1±0.9 on the accepted olmocr benchmark, outperforming general vision models like GPT-4o and Gemini 2.5 Flash on strict document parsing accuracy.
Flexible Dual Inference Topologies
Supports lightweight remote batch processing via an OpenAI-compatible vLLM server, or completely local, standalone script tracking using the HuggingFace transformers library.
Robust 90+ Multilingual Script Support
Features native recognition matrices covering Latin, Cyrillic, and right-to-left scripts (Arabic/Persian), with major performance enhancements across South and Southeast Asian languages.
Native Image & Caption Harvesting
Detects embedded photographs, line charts, and chemical diagrams, extracting the raw asset files while cataloging their exact semantic captions natively into the metadata stream.
Streamlit Web Viewport Integration
Includes a built-in graphical workspace tool triggered via chandra_app, allowing developers to upload documents and visually audit parsed text arrays side-by-side.
Eliminates Document Structure Scrambling
Maintains authentic visual reading patterns, completely preventing the text fragmentation and out-of-order column clipping errors that break old engines like Tesseract.
Generates Clean Downstream AI Assets
Converting sheets natively into valid Markdown headings (`##`) and structured math equations ($$) creates perfect, clean assets for LLM reasoning ingestion.
Free Core Layer for Small Business & Startups
The open-weight framework is fully accessible and free for research, personal use, and early-stage companies under $2M in funding or ongoing revenue metrics.
Cons
Restrictive Top-Tier Commercial Terms
The custom OpenRAIL-M weight license prevents large corporations or enterprise networks exceeding the $2M boundary from deploying it competitively without custom API licensing.
Noticeable Hardware Requirements
Running the fully layout-aware 5B parameter structure locally for fast, production-grade throughput scales optimally when paired with solid modern NVIDIA CUDA hardware.
