Nace.AI Launches Drex for Document Processing: Small Decision Models That Check OCR Output Page by Page
The new engine pairs layout-aware parsing with a decision model that verifies and corrects only the parts that are wrong, claiming a 0.834 Parse Index and corrections 28–38% faster than full regeneration.
Nace.AI has introduced Drex for Document Processing, an intelligent document processing engine built around a Read, Check and Correct pipeline. Instead of using a generative model to rewrite every page, the engine first produces layout-aware Markdown, then a small Drex decision model compares the extraction against the original page image, flagging mismatched text, broken tables or checkboxes lost in scanning. An agentic correction step then repairs only the affected fields, and if a correction comes back empty or implausible the original output is kept.
The company reports a Parse Index of 0.834 across twelve parsers tested, an OmniDocBench score of 95.7 and 91.9 on smartphone-photo documents, with corrections running 28 to 38 percent faster than full-page regeneration. The service supports 30 file types, is priced at $7 per 1,000 pages and is offered as managed cloud, dedicated cluster, customer-owned AWS or Azure, and fully on-premises deployment.
All benchmark figures are vendor-reported and have not been independently replicated. The architecture nevertheless signals a broader shift in document AI, where cheap parsers handle reading while low-cost decision models act as per-page verifiers, reducing silent extraction errors before data reaches downstream systems.