Skyward’s Open Source RAG APP ‘SWParse’ Surpasses Claude in Document Parsing!

By [Ksenia Kapoor](/content/author/kkapoor "link-to-author"/index.html)
February 25, 2026
3 min read

A breakthrough for specialized open-source tools in the AI ecosystem

Text similarity measurements comparing ground truth to various processing methods

At Skyward, we’ve been tackling one of the most persistent challenges in the RAG ecosystem: transforming diverse document formats into high-quality, structured inputs for language models. Today, we’re excited to share that our recently open-sourced solution, SWParse, has achieved performance surpassing that of Claude in document parsing accuracy!

Open Source stays winning!

When benchmarking specialized tools against general AI systems, we typically expect a significant performance gap. However, our testing revealed that SWParse is achieving slightly superior results compared to sophisticated commercial AI systems.

We’ve rigorously benchmarked SWParse using standardized metrics from the Omni OCR Benchmark. Our results show SWParse achieving 70.4% JSON accuracy on 923 test documents, which edges out Claude 3.5 v2’s 70.0% accuracy over the same dataset!

Benchmark results comparing JSON accuracy across different processing pipelines

In our comprehensive testing:

  • SWParse achieved 70.4% JSON accuracy across 923 diverse documents, comparable to Claude’s 70.0% accuracy across 1000 images
  • In text similarity evaluation, SWParse demonstrated 50.9% similarity to ground truth, significantly outperforming Claude which only achieved a 38.7% similarity — showing SWParse’s superior ability to maintain textual fidelity!

These performance metrics follow the rigorous methodology established by the Omni OCR Benchmark, ensuring standardized comparison against leading solutions in the field.

The Critical Link in RAG Systems

Document parsing quality directly impacts retrieval-augmented generation effectiveness. Even the most advanced language models struggle when fed poorly structured document representations. SWParse addresses this vital component in the RAG chain, focusing exclusively on optimizing document processing before they reach the language model.

Specialized Tools for Specialized Tasks

While Claude excels as a general-purpose AI assistant, SWParse demonstrates the value of purpose-built tools designed for specific tasks in the AI pipeline. Our system is architected specifically for document parsing:

  • Document structure detection preserving semantic relationships
  • Table extraction with custom declarative syntax
  • OCR for images with layout preservation using Surya
  • Async processing with configurable workers for scalability
  • Containerized design for easy deployment
  • API-first architecture for integration flexibility
  • Output formats including Markdown, JSON, CSV, HTML, and plain text

The Open Source Advantage

Perhaps most significantly, SWParse achieves this performance as an open-source solution. This means:

  1. Transparency: The entire processing pipeline is visible and auditable
  2. Community improvement: Contributions can accelerate development beyond what closed systems allow
  3. Customizability: Organizations can tailor the tool to their specific document types and needs
  4. Cost efficiency: No API usage fees or pricing tiers

Continuing the Journey

While we’re proud of SWParse’s current performance, we see this as just the beginning. The open-source model allows for rapid iteration and improvement, with several enhancements already in development:

  • PPTX conversion support
  • Advanced caching systems for improved performance
  • Enhanced table extraction from PDFs and images
  • User query-based extractions
  • Addressing failure cases to improve overall reliability

We’re actively working on resolving the issues with the 77 documents where processing failed in our benchmark tests. Our team is dedicated to fixing these edge cases, and we’re eager to publish full benchmark results across the complete 1000-image test set. We believe these improvements will further strengthen SWParse’s competitive position against commercial solutions.

Join the Open Source RAG Community

We’re releasing SWParse on GitHub today and welcome contributions, feedback, and feature requests from the community. Check out our repository at https://github.com/skywarditsolutions/swparse to get started.

As RAG applications become increasingly central to AI deployment, having open-source, high-performance document parsing represents a crucial component in democratizing this technology. We believe SWParse demonstrates that specialized open-source tools can approach the capabilities of commercial AI systems in their areas of focus.