The RIKEN Center for Computational Science has started using the GFLOPS AI assistant "AskDona" to provide a real-time response service utilizing generative AI technology to users of the Fugaku support site.

Photo: From left, RIKEN Shoji, Head of the Operations Technology Department, Mikami, Senior Technical Staff, Nakamura, Engineer, GFLOPS President Morimoto, Co-Representative Suzuki

We are pleased to announce that GFLOPS, Inc. (hereinafter referred to as GFLOPS) and JAFCO Group, Inc. (hereinafter referred to as JAFCO) have implemented GFLOPS' generative AI assistant "AskDona"*1 on the support site for the supercomputer "Fugaku" (hereinafter referred to as "Fugaku") to provide rapid and accurate responses to inquiries regarding the use of Fugaku.

Background to the Introduction of AskDona The Fugaku support site provides technical support to users both domestically and internationally. While it is important to comprehensively provide technical information to users, the vast amount of information can make it difficult for users to find necessary information among numerous technical documents created in various formats to solve problems. Promoting self-resolution of issues by users has become a recent challenge of the Fugaku support site through the provision of high-function search services and high-precision automated response services to questions. Against this backdrop, discussions began on the introduction of AskDona as a part of JAFCO's entrepreneurial support program "First Leap"*2, in which RIKEN and GFLOPS engaged in technical discussions regarding generative AI. After a verification and evaluation process lasting about six months, the solution using AskDona was highly evaluated and its implementation was decided.Purpose of the Introduction of AskDona Aiming to enhance the FAQ page and other sections of the Fugaku support site, we will leverage the effectiveness of the question-answering system using rapidly developing generative AI technology and realize a real-time service that provides answers in a chat service format to users. The answers will be optimized to meet the needs of each Fugaku user by applying the learning of the manual and related materials of Fugaku using a technique called retrieval-augmented generation (RAG)*3 that underpins AskDona.

Reasons for Adopting AskDona

1) Answer Accuracy: Clear the review criteria with an answer accuracy of over 80% utilizing retrieval-augmented generation (RAG) technology (Patent pending*4) We constructed a RAG using not only the over 16,000 pages of PDF manuals for Fugaku but also HTML format user guides and FAQ pages as sources of answer information. GFLOPS meets the technical review requirements set by RIKEN for answer accuracy using a system that achieves high answer accuracy from large-volume data that is currently patent pending.

2) User Usability: Provides solutions that include the flow of knowledge search through the application of techniques using retrieval-augmented generation (RAG) In this case, in addition to providing an interactive chat format through RAG, a document search function will also be implemented. The document search function will use the mechanism of RAG but will not generate response texts; instead, it will provide highly relevant documents in a list format. By integrating the ticket creation function*5 of the existing Fugaku support site with AskDona, AskDona will seamlessly link the history with tickets on the Fugaku support site if human assistance is required. This will greatly enhance user convenience.

Photo: Example of the usage screen of the generative AI assistant: AskDona

‍

Comment from Fumiyoshi Shoji, Head of the Operations Technology Department
RIKEN Center for Computational Science I had thought that utilizing generative AI might dramatically improve user support for "Fugaku" when I was introduced to GFLOPS by JAFCO and began consideration. It immediately became clear that the AI assistant AskDona developed by GFLOPS was more effective than I had imagined, but even more impressive was the speed at which they responded to our requests. I hope to continue collaborating to enhance the services for "Fugaku" users in the future.

Comment from Maria Morimoto, President of GFLOPS, Inc. Since the founding, we have been engaged in research on RAG as a technology that makes it easy for businesses to utilize generative AI, and with the support of JAFCO's extensive network in customer development, we are very honored that RIKEN has adopted AskDona. The opportunity to utilize AskDona with feedback from RIKEN at the Fugaku support site, which is used by people all over the world, is extremely valuable, and I am confident that it will significantly contribute to promoting the practical use of generative AI application technology and enhancing user satisfaction.

About AskDona AskDona is a generative AI assistant that centrally handles conversations through SMS, SNS, Web, and business chat. It learns internal knowledge and automatically generates answers. In addition, it integrates more than ten task execution tools that can link with business applications and autonomously carry out tasks based on instructions. The core technology of AskDona utilizes a RAG (Retrieval-Augmented Generation) system employing cutting-edge large-scale language models such as ChatGPT, Claude, and Gemini. A patent application is currently pending for the high answer accuracy technology of RAG and the system for simultaneously connecting to RAG from multiple channels.‍The URL for the AskDona site: https://askdona.com‍

About GFLOPS, Inc. Founded by two former Google employees with support from the JAFCO Group's entrepreneurial support program "First Leap," GFLOPS aims for a hybrid work model between AI and humans, providing solutions that utilize generative AI application technology. Company name: GFLOPS, Inc. (English name: GFLOPS Co., Ltd.) Representative: Maria Morimoto, Co-representative: Ryo Suzuki, Head office: Shibuya, Tokyo, Business: Development and provision of AI services utilizing large-scale language model (LLM) generative AI technology, Web: https://gflops-ai.com/‍

About JAFCO Group, Inc. Since its founding in 1973, JAFCO has worked with stakeholders to utilize new technologies to tackle challenges beyond existing constraints and common sense. With the motto "Creating a cycle of growth through investment in challenges," we are committed to creating innovative technologies and services around the world. Company name: JAFCO Group, Inc. / English name: JAFCO Group Co., Ltd. President: Keisuke Miyoshi Head office location: 24th Floor, Toranomon Hills Mori Tower, 1-23-1 Toranomon, Minato Ward, Tokyo, Web: https://www.jafco.co.jp/‍

About RIKEN As the only comprehensive research institute for natural sciences in Japan, RIKEN conducts research in a wide range of fields including physics, engineering, chemistry, mathematics and information science, computational science, biology, and medical science, and is actively involved in joint and commissioned research with universities and companies to disseminate research results to society, as well as proactively promoting technology transfer of intellectual property to the industrial sector. Director: Masataka Gohama, Location: 2-1 Hirosawa, Wako City, Saitama Prefecture, Web: https://www.riken.jp/‍

*1 AskDona: A generative AI assistant that centrally handles conversations through SMS, SNS, Web, and business chat. It learns internal knowledge and automatically generates answers. It integrates more than ten task execution tools that can connect with business applications and autonomously perform tasks based on instructions. The core technology of AskDona employs a RAG (Retrieval-Augmented Generation) system using cutting-edge large-scale language models such as ChatGPT, Claude, and Gemini. *2 First Leap: The name of the support program aimed at entrepreneurship backed by JAFCO, which boasts the largest cumulative number of invested startups in Japan. *3 Retrieval-Augmented Generation (RAG): The technique that denotes the combination of information retrieval (retrieval) and generation (generation) by providing external information to a large-scale language model.

*4 Patent pending: Applications are pending for technology to improve answer accuracy of retrieval-augmented generation (RAG). *5 Ticket creation function: A function where users of Fugaku can create tickets by submitting inquiry contents through a form, enabling the inquiry handler to respond.