Ur AI Launches 'Nebula' to Convert Japanese Business Documents into AI-Ready Format

Ur AI (Headquarters: Tokyo, CEO: Ella Sandeep) announced on July 13, 2026, the launch of 'Nebula,' a document AI product that converts Japanese business documents into AI-ready formats. Nebula is offered in two forms: 'Nebula Frontier,' available via platform and API, and 'Nebula Sovereign,' a fully self-hosted version that operates within the customer's environment. Nebula Frontier addresses challenges such as difficulty in feeding PDFs and other documents into AI applications or generative AI tools, or ballooning token costs with each document submission. Nebula Sovereign responds to cost and security constraints arising from expanding AI adoption.

Concurrently, Ur AI released a new benchmark, 'RCRR Benchmark,' designed to measure how well the 'meaning' of documents is preserved after conversion, and published comparative evaluation results for multiple document AI products and frontier VLM APIs, including Nebula. Nebula Frontier achieved a total RCRR score of 94.4, the highest among evaluated document AI products. It is statistically equivalent (no significant difference) to leading frontier VLM APIs such as Fable 5, GPT-5.6 Sol, and Gemini 3.1 Pro, and recorded the highest score (94.3) among all evaluated systems in text and table questions. Nebula Sovereign scored 87.3, statistically equivalent (no significant difference) to Microsoft Azure Document Intelligence.

Background of the Announcement

Much of corporate knowledge is accumulated in PDFs, scanned documents, and slides. To enable generative AI to read these documents, conversion into machine-readable formats is essential. However, during this process, relationships between numbers and items, table structures, and chart meanings can be lost.

Even if characters are correctly recognized, AI may provide incorrect answers if the document structure is corrupted. Ur AI identified this challenge as an issue of 'meaning preservation in document conversion' and developed the RCRR Benchmark.

Overview of the RCRR Benchmark

RCRR (Reading-Comprehension Recovery Rate) is a metric that measures not character matching accuracy, but how well an AI, after reading the converted document, can answer questions that a human could answer by referring to the original page.

This benchmark was constructed using 99 pages of real Japanese IR materials published on TDnet, with 1,410 independently verified questions administered and scored under identical conditions across all systems. Since real-world operations involve handling PDFs exceeding 100 pages or multiple files, frontier VLM APIs were evaluated using page image inputs (real-world conditions).

Evaluation Results

The evaluation results show that Nebula Frontier achieved a total RCRR score of 94.4, the highest among evaluated document AI products. Among frontier VLM APIs, Fable 5 scored 94.6, GPT-5.6 Sol scored 94.0, and Gemini 3.1 Pro scored 93.7, all statistically equivalent (no significant difference) to Nebula Frontier. Microsoft Azure Document Intelligence scored 88.2, Nebula Sovereign scored 87.3, and Mistral OCR scored 73.6. Other commercial and open-source OCR systems scored between 20.2 and 85.9, with full system comparisons detailed in the technical report.

Figure 1: Comparison of RCRR Total Scores for Document AI Products and Frontier VLM APIs

Nebula Frontier is an orchestration version that runs frontier VLM APIs through Ur AI's conversion pipeline (full-page conversion, combined use of PDF text layer). In this evaluation, it statistically significantly outperformed Microsoft Azure Document Intelligence, the most powerful commercial cloud OCR, by 6.2 points. It is statistically equivalent (no significant difference) to the latest frontier VLM APIs (Fable 5, GPT-5.6 Sol, Gemini 3.1 Pro) and recorded the highest score of 94.3 among all evaluated systems in text and table questions.

Nebula Sovereign is a fully self-hosted, fine-tuned version where a fine-tuned Qwen3-VL-32B operates exclusively on GPUs managed by the customer. It is statistically equivalent (no significant difference) to Azure Document Intelligence (88.2) and outperformed Mistral OCR (73.6) by approximately 14 points.

For chart-related questions specifically, Nebula Frontier scored 94.7, Nebula Sovereign scored 77.3, both surpassing Azure Document Intelligence's 69.1. In documents containing charts, even if numerical values are correctly read, the correspondence between values and items may be misinterpreted. RCRR is characterized by its ability to evaluate such cases where 'characters are readable but meaning is corrupted.'

Figure 2: RCRR Score Comparison for Chart Questions

Figure 3 shows an example of a chart document targeted by the RCRR Benchmark. In such waterfall charts showing profit and loss changes, even though each number is clearly displayed on the page, the correspondence between 'which number corresponds to which item' may collapse during conversion. For example, if the correspondence between items such as foreign exchange differences, price/MIX, and fixed costs and their values is incorrect, an AI reading the converted document may plausibly answer with another actual number present on the page. RCRR is a benchmark designed to evaluate such cases where 'characters are readable but the semantic correspondence is broken.'

Figure 3: Example of a Chart Document Targeted by the RCRR Benchmark

Nebula Delivery Models

Nebula Frontier: An orchestration version that integrates frontier VLM APIs. Available via platform and API, with availability starting today. Platform: nebula.ur-ai.net, API documentation: ocr.ur-ai.net/docs.

Nebula Sovereign: A fully self-hosted, fine-tuned version operating on GPUs managed by the customer. Since data does not leave the customer's environment, it is offered to enterprises handling highly sensitive documents in finance, insurance, semiconductors, and healthcare.

Pricing

Nebula Frontier: Starting at 10 yen per page (excluding tax, pay-as-you-go). PDFs under 100 pages can be tested free of charge on the platform.

Nebula Sovereign: Individual quotes based on deployment model. Fixed-cost operation is possible with data remaining within the customer's environment.

About the Self-Hosted 'Nebula Sovereign'

Nebula Sovereign, delivered as a self-hosted solution operating within the customer's environment, eliminates the risk of data leakage outside the organization. Its adoption is anticipated in industries such as financial institutions and healthcare providers that handle extremely sensitive documents.