Accuracy Overview
When evaluating document intelligence tools, accuracy is usually the first question:- How do you compare to other OCR tools and/or LLM-based approaches?
- Have you run benchmarks?
Benchmarks & Results
We’ve evaluated Cardinal on a range of public benchmarks as well as internal test suites.- Full results: Email us at team@trycardinal.ai to get access!
- Performance: Cardinal consistently matches or outperforms other tools and generic LLMs on dense tables, structured forms, and real-world edge cases.
- Strengths: excels at complex tables, small text, mixed layouts, and annotations that typically break baseline LLMs.
- Limitations: as with any system, extremely degraded scans, messy handwriting, or corrupted PDFs may still require human review.
Why Not Just an LLM?
It’s tempting to just “give the PDF to GPT/Gemini” and hope for structured output. The problem is:- LLMs hallucinate — they may invent rows/columns or misalign values.
- LLMs lose structure — PDFs with tables, checkmarks, barcodes, or multiple columns rarely survive a naive LLM parse.
- Cardinal preserves fidelity — we give you bounding boxes, cropped images, and schema-mapped JSON so you know where data came from.
Cardinal is not “just an LLM wrapper.” We combine OCR, layout models, and post-processing layers with selective LLM use. This hybrid approach gives deterministic structure with AI flexibility where it matters.