Ur AI Launches Nebula Document AI with New Meaning-Preservation Benchmark
The RCRR Benchmark provides a standardized way to evaluate how well document conversion preserves meaning for AI consumption, a problem that has led to incorrect AI responses when structure is lost.
Reporting from 1 source: ASCII.jp.
Ur AI announced Nebula, a document AI product that converts Japanese business documents into AI-readable formats, available as a platform/API version (Nebula Frontier) and a self-hosted version (Nebula Sovereign). The company also released the RCRR Benchmark to measure meaning preservation after conversion. Nebula Frontier scored 94.4, the highest among document AI products and statistically equivalent to frontier VLMs like Fable 5 and GPT-5.6 Sol.
Ur AI's new benchmark, the RCRR, is built from 99 pages of Japanese IR materials and 1,410 independently verified questions. It measures how well an AI can answer questions from a converted document that a human could answer from the original page, rather than character match rates. The company argues that standard OCR often breaks table structures and chart meanings, leading to incorrect generative AI responses.
Nebula comes in two forms. Nebula Frontier runs on Ur AI's conversion pipeline using frontier VLMs and scored 94.4 on the RCRR, the highest among document AI products and statistically equivalent to Fable 5 (94.6), GPT-5.6 Sol (94.0), and Gemini 3.1 Pro (93.7). Nebula Sovereign, a fully self-hosted version, scored 87.3, equivalent to Microsoft Azure Document Intelligence. Mistral OCR scored 73.6, and other OCRs ranged from 20.2 to 85.9.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.