Release of the operational report on the implementation and operation of the generative AI chat "AskDona" for user support of the RIKEN Supercomputer "Fugaku".

GFLOPS Co., Ltd. (GFLOPS) and the RIKEN Center for Computational Science (R-CCS) are pleased to announce the release of a report summarizing the implementation and operational results of the generative AI chatbot “AskDona” on the Fugaku supercomputer support site.

◾️ View the Report
The full report is available at the following URL:
https://askdona.com/askdona-riken-report/

◾️ Overview of the Report

Generative AI Chatbot “AskDona” Established as Users’ “Thinking Partner”

“AskDona” has become firmly established as a daily support tool, handling an average of approximately 680 queries per month. In the 10 months following implementation, the proportion of “complex queries”—those requiring the integration of multiple information sources—increased approximately 4.3 times, from 2.3% to 10.0%. This indicates users have begun recognizing generative AI not merely as a “search tool,” but as a sophisticated “thinking partner” to assist with advanced problem-solving.

Achieved Up to 61% Efficiency Improvement in Human Support at the Fugaku Support Site

By consolidating inquiries through “AskDona,” there has been a dramatic reduction in tickets issued for human support, achieving a maximum 61% decrease in April 2025 compared to the same month in the previous year. Instances where issues previously requiring approximately four hours were resolved within approximately five seconds have been observed. This improvement significantly contributes to 24-hour prompt support availability and allows support personnel to focus on high-value-added tasks.

Demonstrated Superior Performance with Proprietary Agentic RAG

The accuracy of responses to “complex queries” was benchmarked against major cloud-based RAG services (Azure, GCP, AWS) and the OSS framework (LangChain). AskDona recorded an average score of 83 points, surpassing the comparative systems’ average score of 61 points by 22 points. Significant performance advantages were observed in areas requiring multi-source information integration (completeness) and specialized practicality, clearly demonstrating its technical superiority.

By publicly sharing the insights obtained through this report, GFLOPS and R-CCS aim to further promote the social implementation of generative AI technology.

About GFLOPS Co., Ltd.
GFLOPS Co., Ltd. specializes in unique solutions combining Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) technologies, providing AI solutions that enhance operational efficiency and foster innovation in enterprises.

Company Name: GFLOPS Co., Ltd.
Founder: Maria Morimoto, Co-representative: Ryosuke Suzuki
Headquarters: Shibuya Tokyo, Japan
Business: Development and provision of AI services utilizing Large Language Models (LLMs) and generative AI technologies.
Website: https://gflops-ai.com/

About RIKEN Center for Computational Science (R-CCS)
The RIKEN Center for Computational Science (R-CCS), part of Japan’s only comprehensive natural sciences institute, RIKEN, aims to serve as a world-leading research center and a core institution in Japan for computational science. R-CCS is involved in research and development for next-generation supercomputers following Fugaku, as well as collaborations in emerging computational technologies such as quantum computing. It also actively develops foundational models for rapidly advancing “AI for Science,” with results expected to significantly benefit everyday life, industry, and economic growth.

Director: Satoshi Matsuoka
Headquarters: Kobe, Hyogo, Japan
Website: https://www.riken.jp/research/labs/r-ccs/index.html

※1 AskDona: Generative AI chatbot implemented on the Fugaku support site by R-CCS. Employs Retrieval-Augmented Generation (RAG) technology to answer user queries by referencing Fugaku’s manuals and technical documents.
URL: https://askdona.com

※2 Complex Queries: Queries (user messages or questions) that require cross-referencing multiple documents and integrating information from multiple sources for comprehensive analysis and appropriate responses.

※3 Inquiry Ticket: Individual inquiries submitted by Fugaku users to support staff via an inquiry system, managed individually as “tickets.”

※4 RAG (Retrieval-Augmented Generation): A technique whereby large language models generate answers by referencing external documents or data sources in real-time, significantly reducing hallucination risks and improving answer accuracy.