~ Immediate access to vast amounts of internal knowledge improves the comprehensiveness of safety management. To be implemented in five domestic factories by the end of the year. ~

Our company will begin full-scale development in April 2026 of a new system that incorporates Cinnamon AI's "Super RAG™" into the KY (Hazard Prediction) suggestion system (hereinafter "MGC-KYAS") currently in operation at five domestic plants of Mitsubishi Gas Chemical Company, Inc. (Head office: Chiyoda-ku, Tokyo; President and CEO: Yoshinori Isahaya; hereinafter "Mitsubishi Gas Chemical"), which will enable more accurate in-house knowledge search and response generation. The system is scheduled to be introduced sequentially within the year.

MGC-KYAS has a history of near-miss activities*¹This system effectively supports pre-work risk assessment (KY) activities by extracting case studies and information related to on-site work from a vast database of case studies obtained through this process. Based on Cinnamon AI's advanced AI solutions, this system has been introduced to Mitsubishi Gas Chemical's five domestic plants since around June 2023.

The introduction of "Super RAG" will enable high-precision reading of vast amounts of documents stored in various formats across factories and departments. This knowledge can then be used for a variety of purposes, including immediate reference to similar past accidents, optimization of work procedures, and creation of safety training materials. Furthermore, by utilizing AI in hazard prediction activities, we can prevent biases in perspective and stagnation that can easily occur when relying solely on humans, significantly improving the comprehensiveness of hazard prediction. In the face of a worsening labor shortage, we will build a system that reliably transmits the valuable hazard information know-how held by veterans via AI, further strengthening the prevention of human error and enhancing the awareness of workers.

*1 An initiative to report and share incidents that did not result in accidents or disasters, but where employees had a "near miss" or "quick stop," in order to use this information to inform organizational disaster prevention activities.

■ Background: The challenges of passing on the knowledge of veterans and the formalization of risk assessment activities.

　
In manufacturing, "KY (Hazard Prediction) activities," which nip potential accidents and disasters in the bud before work begins, are essential for safety management. However, in recent years, as automation of equipment has advanced, opportunities for employees to directly engage in safety and disaster prevention work on-site have decreased, raising concerns about a decline in workplace risk awareness. Furthermore, although a vast amount of near-miss incidents and accident/work-related injury data has been accumulated, it is difficult to utilize effectively because it is stored in different formats from site to site. This has led to challenges such as activities becoming mere formalities and requiring a great deal of time to extract necessary information.

To address these challenges, Cinnamon AI has collaborated with Mitsubishi Gas Chemical to develop "MGC-KYAS," a KY suggestion system that accurately and automatically suggests potential hazards in operations based on a vast amount of past knowledge. We have been promoting its implementation at five of our domestic plants.

■ Main achievements and features of the "Super RAG" implementation

　
By implementing Cinnamon AI's proprietary RAG (Retrieval-Augmented Generation) technology, "Super RAG," into "MGC-KYAS," the following three points will be achieved:

1. Improving the quality of KY (Kiken Yochi - Hazard Prediction) activities through rapid information extraction.

Traditionally, researching similar past cases required searching through a vast number of files. However, with this system, simply entering, for example, an equipment number or task name allows the AI to instantly extract highly relevant cases. By automatically presenting multiple "hazards hidden in the work," which is the first step in hazard prediction training, it minimizes the burden of information gathering on-site and significantly improves the quality of KY activities.

2. Reducing operational burden through direct use of unstructured data

Thanks to the advanced analytical capabilities of "Super RAG," it is possible to generate highly accurate answers directly from original files such as PDFs, without the need to process them into structured data in Excel or other programs. This significantly reduces the workload of data registration on-site, while providing high convenience by making hidden knowledge immediately available for use.

3. Passing on veteran expertise with AI, and breaking the monotony of risk assessment activities.

By having AI instantly retrieve the tacit knowledge of veterans, safety know-how, which tends to be highly individualized, can be standardized, ensuring reliable knowledge transfer even amidst labor shortages. The function that allows the AI to clearly indicate the basis of its answers (referenced documents) enables younger employees to directly learn the judgment criteria and past lessons of experts, leading to improved risk awareness.

Furthermore, by having AI present "dangers lurking in work" from a new perspective, we can break the monotony of thinking that often occurs in daily work, prevent past serious accidents from being forgotten, and continuously update the company-wide safety culture.

■ Regarding the future

Cinnamon AI will strengthen its collaboration with Mitsubishi Gas Chemical and explore expanding the scope of application of "MGC-KYAS." Furthermore, with an eye on deploying it to group companies and partner companies, we will promote the advancement of safety management in a wider range of areas using AI and the realization of sustainable manufacturing sites.

■ "Super RAG" has an answer accuracy of over 90% without customization^*2

Cinnamon AI's "Super RAG" is a unique AI product that enables the creation of an LLM environment without tuning, capable of analyzing complex documents such as tables and figures and generating highly accurate answers. Its greatest feature is that it goes beyond simply enhancing some functions, instead "optimizing the entire process"—preprocessing, searching, and generation—in a general-purpose and highly accurate manner. In particular, in preprocessing, unlike conventional RAGs that mechanically divide by the number of characters, it uses its unique layout analysis technology to automatically identify titles and figures/tables, and optimally chunks the document while preserving its meaning and structure. This highly structures unstructured data that is unsuitable for data utilization, enabling extremely high-precision automation of business processes using RAG systems.

"Super RAG" combines vector search and graph structure search to derive accurate answers without overlooking anything, even covering relationships not explicitly stated in documents. No prior data processing is required to optimize the search; simply register specialized information such as financial statements, technical documents, and internal regulations, and you can instantly create an AI assistant on-site with no-code operations in as little as five minutes. Its versatility allows it to be deployed to multiple use cases in a single environment, instantly providing "business-ready accuracy" in a wide range of areas, including responding to inquiries, creating reports, and small-lot, high-mix operations that require tacit knowledge.

*2 The accuracy of the answer may vary depending on the type of document.

https://cinnamon.ai/super-rag/

▼ Examples of how to use "Super RAG"

[Human Resources/Legal Department]Answering work regulations and checking contracts
[Sales Department]Product information confirmation, FAQ automation
[Manufacturing site]Knowledge sharing and safety management support
[Technology/IT department]Development knowledge search, business system support
[Management and Planning Department]IR information responses, management information search
*The assistant function also allows department staff to build and deploy the system.

Complex documents including figures and tables that can be used with "Super RAG" (image)

* "Super RAG™" as used in this press release is a trademark of Cinnamon Co., Ltd.

■ Mitsubishi Gas Chemical Company Profile

Since its founding, Mitsubishi Gas Chemical has been committed to creating new technologies and value, supporting people's lives through a wide range of business fields, from basic chemicals such as methanol, xylene, and hydrogen peroxide, to functional products such as high-performance engineering plastics, semiconductor packaging materials, and the oxygen absorber "Ageless®". Under its mission of "Creating value that can be shared with society," Mitsubishi Gas Chemical will continue to contribute to the development and harmony of society by providing a wide range of value based on chemistry.