November 6, 2025
Table Wars: Spin or Win?
Benchmarking the Most Reliable Document Parsing API
Startup says it’s best at reading PDFs; commenters yell “ad!” and speed-test it
TLDR: Tensorlake claims its parser beats Azure and AWS by preserving layout and producing usable data. Commenters cry marketing, say rivals like Gemini and DeepSeek were skipped, and mock slow speeds versus Docling—though some still plan to test it for real-world invoices because good document extraction can save serious time and money.
Tensorlake says it cracked the PDF headache with a new document parser that keeps layouts intact and turns messy files into data machines can actually use. They’re bragging 91.7% accuracy and tout two simple-sounding tests: TEDS (basically, “is that table still a table?”) and JSON F1 (“can a robot actually use this info?”). But the comments? Oh, they brought the popcorn.
The top-voted vibe: “Marketing alert.” One critic blasts the write-up for skipping “real leaders” and not testing Google’s Gemini up front, hinting it might win and be cheaper. Another user tried invoices and says Gemini worked best so far—but still missed plenty—so they’ll try Tensorlake anyway. Then came the speed-shaming: one commenter claims open-source Docling finished a file in 20 seconds while Tensorlake spun its loading wheel for 10 minutes. Cue the “loading spinner Olympics” memes and jokes about benchmarks where the host mysteriously wins.
Others want receipts: people asked for head-to-heads with DeepSeek-OCR, PaddleOCR-VL, MinerU 2.5, and even open-source contenders like chandra. One commenter pointed to the updated OmniDocBench leaderboard for a reality check. Verdict: the tech sounds cool, but the community wants neutral tests, faster results, and fewer “we won our own obstacle course” vibes before they crown a PDF king.
Key Points
- •Tensorlake introduces a benchmark emphasizing structural preservation and downstream usability for document parsing.
- •The Tensorlake Document Parsing model reports 91.7% accuracy on enterprise documents, claiming to outperform Azure, AWS Textract, and open-source alternatives.
- •Evaluation uses two metrics: TEDS for structural fidelity and JSON F1 for schema-based extraction reliability.
- •The methodology includes two stages: structural scoring of Markdown/HTML outputs and downstream JSON extraction via GPT-4o with fixed schemas.
- •Public datasets OCRBench v2 and OmniDocBench are used; Tensorlake audited and corrected OCRBench v2 ground truth and published an updated dataset.