구글 클라우드는 포스코홀딩스가 구글 제미나이(Gemini) 1.5 모델을 활용해 생성형 AI 기반 소재 기술/산업 동향 리포팅 시스템 및 지식 검색 Q&A 포털을 개발하고, 구글 클라우드를 기반으로 자사 AI 플랫폼에 데이터부터 머신러닝(ML)까지 유연한 파이프라인을 구축했다고 밝혔다.
Building a news and knowledge search portal based on the Gemini model
POSCO has developed an AI reporting AI and knowledge search Q&A system that reports the latest news trends related to materials using Google's generative artificial intelligence (AI), and has built a search portal that encompasses all internal facility knowledge information.
Google Cloud announced on the 5th that POSCO Holdings developed a generative AI-based material technology/industry trend reporting system and knowledge search Q&A portal using the Google Gemini 1.5 model, and built a flexible pipeline from data to machine learning (ML) on its AI platform based on Google Cloud.
POSCO Applied AI Research team has developed a generative AI-based materials news trend reporting system to quickly identify the latest global trends in the materials field and establish effective business and research strategies.
The entire process is automated, from collecting and translating multilingual news to recommending news based on ranking algorithms, summarizing the text, and creating and sending email reports on daily and weekly issues on technology/industry by country.
The research team went one step further than just reporting on material news trends and expanded the knowledge search and Q&A system that synthesizes the latest material news and related documents that are continuously accumulating. When users ask questions about their personalized material interests in natural language, they can receive the information and insights they want with over 90% search and response accuracy. Currently, more than 500 employees from around 10 affiliates within the POSCO Group, including POSCO Holdings, are using the system every day.
POSCO Holdings operates both a generative AI-based material news trend reporting and knowledge search Q&A system using Google Cloud, and selected Gemini 1.5 Pro as the generative AI’s large-scale language model (LLM) engine.
In particular, the research team confirmed that when the Gemini model was combined with the high-performance RAG technology to build a proprietary knowledge AI agent system for POSCO Group based on its own high-performance augmented search generation (RAG) architecture and document processing AI, it showed better performance and efficiency than third-party LLMs. In addition, by utilizing the Google Cloud Platform, time and cost were significantly reduced.
POSCO Holdings, in collaboration with POSCO, recently expanded and built a Q&A system for searching knowledge related to facility electricity. First, it processed over 30,000 pages of POSCO-accumulated facility electricity-related documents using multi-modal AI specialized for POSCO documents, and built a knowledge vector database based on this.
Afterwards, we built a system that combined our own high-performance RAG technology and Gemini model. Users can easily and effectively search for information and obtain answers to questions about equipment needed for work with an accuracy of approximately 90%. The pipeline of the AI system is operated based on Google Cloud and is integrated with knowledge services utilized by hundreds of employees daily and thousands of employees monthly within POSCO, contributing significantly to improving work productivity.
In addition, POSCO Holdings highly evaluated Google Cloud's flexible service capabilities that provide the most optimized environment for complex internal data and self-developed algorithms, and built the overall data pipeline of the in-house Applied AI platform consisting of multiple clouds with Google Cloud. The construction was a joint effort between Google Cloud and AI Premier Partner TPCG (Tipping Point Consulting Group), and currently, various data and AI/ML solutions of Google Cloud, such as Data Catalog, BigQuery, Vertex AI, and Apigee, are being organically utilized.
POSCO Holdings is conducting research and development on multimodal generative AI and AI-based simulation to accelerate material development, and recently announced the highest level of stability results in the crystal material structure generation sector using LLM. In the future, we plan to consider additional introduction of cloud environments for innovation through AI agent systems and laboratory autonomy.
“POSCO Group has been investing in AI application research for many years and has recently been working on developing and applying generative AI application algorithms,” said Woosang Lim, a research fellow (executive director) at POSCO Holdings Future Technology Research Institute, who presented these technologies and results at the Google Cloud Summit Seoul 2024. “In order to properly utilize the self-developed AI model, it was essential to build an infrastructure that could effectively serve data and models,” he said. “Thanks to Google Cloud’s outstanding AI capabilities, abundant experience, and close support from engineers, we were able to quickly and efficiently establish a foundation for generative AI innovation.”
Ji Ki-sung, President of Google Cloud Korea, emphasized, “Our collaboration with POSCO Group, a leader in future materials and energy, is an example of how the adoption of generative AI is increasing in the manufacturing industry,” and “We are honored that Google Cloud’s powerful generative AI technology and AI-optimized infrastructure can support POSCO Group’s R&D and business innovation.”