Home > OCR > Chandra OCR Resources: GitHub, Demo, and HuggingFace Guide

Chandra OCR Resources: GitHub, Demo, and HuggingFace Guide

by Rachel Jones 2026-03-03

Chandra OCR brings a new level of accuracy to reading complex documents. It handles everything from handwritten notes and messy forms to tables, math equations, and multi-column layouts, extracting text, images, and data.

In this article, you’ll discover how to install Chandra OCR, run it locally or via chandra-o cr huggingface, and test it on your own documents. Step by step, we’ll explore its layout-aware features and structured outputs. This helps developers, researchers, and businesses work smarter with PDFs and other document formats.

Part 1. What Is Chandra OCR

A Simple Definition of Chandra OCR

Chandra OCR is an open-source, layout-aware optical character recognition (OCR) model designed to read and understand complex documents. Unlike traditional OCR tools that only recognize characters, Chandra OCR captures the structure of documents, including tables, forms, handwritten notes, math equations, and multi-column layouts.

It outputs content in structured formats like Markdown, HTML, or JSON, making it ideal for developers, researchers, and businesses needing accurate and organized data from PDFs or scanned files.

Core Features That Set Chandra OCR Apart

• Structured Output & Layout Awareness: Every text block, table, and image is identified and positioned correctly.
• Multilingual Recognition: Supports more than 40 languages.
• Advanced Document Understanding: Not limited to text; it extracts tables, forms, math equations, and captions.
• Flexible Output Formats: Easily export results in HTML, Markdown, or JSON for integration with other tools.

Chandra OCR can be run locally using HuggingFace Transformers, deployed via a vLLM server, or tested interactively with its web app. Its combination of speed, accuracy, and structure-aware output makes it a leading choice in modern OCR solutions.

Part 2. How Chandra OCR Works

How Traditional OCR Works

Traditional OCR tools follow a simple workflow: scanning an image or PDF → recognizing characters → producing plain text output. While this works for simple documents, it struggles with complex layouts, handwritten forms, tables, or multi-column pages. Text often loses structure, making it harder to analyze or reuse.

How Chandra OCR Improves on Traditional OCR

Chandra OCR uses a layout-aware approach that recognizes the position and type of every element on the page. Its workflow includes:

• Layout Recognition: Identifies blocks of text, tables, images, and forms.
• Block Extraction: Processes each block while preserving structure.
• Structured Output: Produces Markdown, HTML, or JSON with full layout information.

This structured output makes documents easier to parse, integrate, or display while keeping tables, forms, checkboxes, images, and math equations intact. Chandra OCR supports over 40 languages, handles handwriting, reconstructs complex forms, and extracts diagrams with captions.

Getting Started

You can run Chandra OCR in Python using the CLI:

Code

pip install chandra-ocr

chandra input.pdf ./output --method hf # Use HuggingFace locally

chandra_vllm # Use vLLM server for batch processing

chandra_app # Interactive web app

Developers can also integrate it into Python scripts with HuggingFace or vLLM, making it easy to automate OCR tasks on PDFs and images. Its advanced layout and structured outputs give Chandra OCR a clear edge over traditional OCR solutions.

Part 3. Chandra OCR Benchmark and Performance

Benchmark Overview

Chandra OCR has been evaluated on independent third‑party benchmarks such as the olmOCR test suite, which measures how well OCR systems handle real‑world document challenges like tables, math, multi‑column text, and handwriting. On this benchmark, Chandra v0.1.0 achieves an overall score of 83.1 ± 0.9, placing it ahead of many other popular OCR systems tested in late 2025.

Benchmark metrics commonly used include:

• Character Error Rate (CER): The percentage of mis‑recognized characters lower is better.
• Word Error Rate (WER): The rate of incorrectly recognized words.
• Layout Accuracy / Structural Fidelity: A measure of how well the system preserves tables, multi‑column text, and document elements.

These indicators reflect not just raw text recognition but how accurately OCR tools maintain structure, a key advantage of Chandra’s layout‑aware design.

How Chandra OCR Performs on Complex Layouts

On task‑specific categories within the olmOCR benchmark, Chandra shows particular strength:

• Table recognition: Scores ~88.0, indicating excellent retention of rows, cells, and headers.

• Math and notation: ~80.3 in older scan math recognition.

• Tiny text (dense content): ~92.3, demonstrating robust extraction of small fonts.

• Handwriting on notes:~90.8, though structured forms with cursive remain a tougher area.

Compared with traditional OCR systems, this performance represents a significant improvement in layout fidelity and structured output because Chandra processes entire pages with spatial context rather than line‑by‑line.

Comparison with Mainstream OCR Tools

Here’s how Chandra stacks up against other prominent OCR systems:

• vs. Traditional OCR (e.g., Tesseract): On clean, single‑column scans, Tesseract remains competitive, but on complex layouts with tables or multiple columns, Chandra reduces layout‑related errors and retains structural order more effectively.
• vs. Modern AI OCR (e.g., GPT‑4o, DeepSeek): On the olmOCR benchmark, Chandra (83.1%) outperforms GPT‑4o’s anchored score (~69.9%) and DeepSeek OCR (~75.4%).
• Real‑world accuracy tests: In practical testing scenarios, Chandra also holds strong word‑level accuracy (~96–98% on clean PDFs) and compares favorably with other models when optimized preprocessing is applied.

Strengths of Chandra OCR

Chandra OCR’s performance shines particularly in areas where traditional or older OCR tools struggle:

• Complex layouts: Superior at extracting multi‑column pages, tables, and forms with structure retained.
• Structured extraction: Outputs in HTML, Markdown, or JSON with layout awareness rather than flat text ideal for automated workflows.
• Multilingual support: Strong recognition across 40+ languages, making it suitable for global document types.
• Handwriting and small text: Performs well on cursive and tiny fonts compared to many alternatives.

Limitations of Chandra OCR

While impressive overall, Chandra OCR has some limitations:

• Hardware requirements: To run efficiently, especially on larger jobs or higher throughput, GPUs are recommended, which might be a barrier for lightweight or budget setups.
• Setup complexity: Compared with simple CPU‑based tools like Tesseract, installing and tuning Chandra can take more effort.
• Handwriting variability: Although strong on many handwritten forms, extremely irregular cursive or stylized scripts may still challenge accuracy metrics.
• Low‑quality scans: Very low resolution or high noise in scanned images can still impact CER/WER scores, necessitating preprocessing steps for best results.

Part 4. How to Install and Use Chandra OCR

Chandra OCR is designed for developers, researchers, and businesses who need high-accuracy OCR with structured outputs. You can run it locally, on Hugging Face, or via a hosted API. Here’s a step-by-step guide.

How to Install Chandra OCR

• System Requirements: Python 3.8+ is required, with a GPU (NVIDIA, CUDA 12.1+ recommended) for optimal performance.
• Virtual Environment: Set up an environment to avoid dependency conflicts.
  

  bash

  python -m venv chandra-env
  
  source chandra-env/bin/activate  # On Windows: chandra-env\Scripts\activate

• Install Package:
  

  bash

  pip install chandra-ocr

• Optional (GPU Optimization): Install Flash Attention and related libraries for better speed.
  

  bash

  pip install vllm transformers accelerate pillow bitsandbytes

• Docker Method (Alternative):


bash

docker pull ghcr.io/datalab-to/chandra-ocr:latest

How to Use Chandra OCR

• Web Interface (Easiest):Launch the Streamlit app to interact with the model directly in your browser.
  

  bash

  chandra_app
  
  # Access it at:
  # http://localhost:8501

• Python SDK:python
  

  python

  from chandra_ocr import ChandraOCR
  
  # Initialize the model
  ocr = ChandraOCR()
  
  # Process a document
  result = ocr.process("path/to/document.pdf", output_format="markdown")
  
  print(result)

• Command Line Interface (CLI):

Specify input and output directories.



bash

chandra-ocr --input_dir /path/to/files --output_dir /path/to/results

Part 5. Chandra OCR vs Other OCR Tools

When choosing an OCR solution, it’s important to compare features, deployment options, and output capabilities. Here’s how Chandra OCR stacks up against popular alternatives like Tesseract, GPT-4 OCR, PaddleOCR, and PDNob OCR:

Please swipe to view

Chandra OCR is ideal for developers who need automation pipelines or want to extract structured data from complex document layouts. Tools like Tesseract or PaddleOCR are also strong for technical OCR tasks, while GPT-4 OCR suits AI-driven data extraction.

However, for everyday PDF management—converting scanned documents into editable files, preserving layouts, and making quick edits—a dedicated PDF editor with built-in OCR, such as PDNob, often provides a simpler and more practical solution. It combines local processing, accurate text recognition, and direct editing, making it user-friendly for both individual and professional workflows.

Part 6. Real-World Use Cases of Chandra OCR

Chandra OCR is not just a research tool it’s highly practical for many real-world applications. Its layout-aware extraction, structured outputs, and multilingual support make it ideal for businesses, academics, and developers. Here are some common use cases:

• Document Digitization: Convert scanned documents, historical records, or messy PDFs into structured text in Markdown, HTML, or JSON. This helps businesses and organizations store, search, and analyze documents efficiently.
• Table Extraction from Reports: Extract complex tables from financial statements, bank statements, invoices, and research reports while maintaining the original structure. Chandra OCR ensures merged cells, formulas, and multi-column layouts are preserved.
• Multilingual Document Processing: With support for over 40 languages, Chandra OCR is perfect for organizations handling global documents. It can recognize text from multilingual PDFs or handwritten forms with high accuracy.
• Academic and Research Documents: Chandra OCR can read textbooks, worksheets, and research papers, including math equations, figures, and tables, making it a great tool for educators, students, and AI researchers.
• AI and Automation Pipelines: Its structured JSON/HTML output is ideal for automation workflows, AI data pipelines, and document intelligence platforms. Developers can integrate Chandra OCR into Python scripts (chandra-ocr python) or use it via Hugging Face (chandra-ocr huggingface) for automated document processing.

Part 7. Common Questions About Chandra OCR

Conclusion:

Chandra OCR is an advanced tool that helps read and extract text, tables, images, and forms from complex documents. It provides structured outputs in JSON, HTML, and Markdown, and is available on chandra-ocr huggingface for developers and researchers. While it excels at automation and handling complex layouts, PDNob is perfect for everyday PDF editing. It provides a fast and easy solution for simple document tasks.

Downloads: 8,678,918

PDNob PDF Editor - Edit, OCR, Work Smarter

rated on Trustpilot >

• Make scanned PDFs searchable and editable with 99% OCR precision
• Batch convert PDFs to Word, Excel, PPT, images, PDF/A, Text, EPUB, etc., up to 30% faster
• Edit PDFs easily like Word, including text, images, watermarks, links, and backgrounds
• Annotate PDF with highlights, comments, shapes, stickers, and stamps
• Run smoothly on any PC without lags or crashes, even on low-spec machines

Free Download

Secure Download

Free Download

Secure Download