KAIST Researchers Unveil 'CONDA' Technology That Maintains Search Accuracy Even as Data Keeps Changing

A research team led by Professor Kim Minsu of the KAIST School of Computing has developed CONDA, a dynamic vector search technology that reliably maintains search accuracy even in environments where data is continuously added and deleted. Generative AI cannot independently learn information that emerges after its training is complete. To address this, Retrieval-Augmented Generation (RAG) — which searches the latest news or internal company documents and uses the results in its answers — is widely used. Currently, many AI services convert the meaning of documents and images into 'vectors,' bundles of numbers, and find materials by linking semantically similar information.

The problem is that data in real companies and on the internet changes constantly. New documents and product information come in while outdated information is revised or deleted. Repeated changes like these can cause 'connectivity collapse,' in which the network of links between data gradually breaks down.

In experiments, CONDA improved search accuracy by up to 24.5% over existing state-of-the-art technologies in environments where data is continuously added and deleted. It also boosted data processing speed by up to 1.9 times. When the research team simultaneously performed searches and data updates for six hours in a large-scale environment containing 100 million data points, CONDA maintained the shortest response times and the highest search accuracy among the compared technologies.

Professor Kim said, "A key value of RAG is that it allows large language models to instantly find and use the latest information they were never trained on, right when it's needed. The significance of this research lies in presenting a data infrastructure that enables AI to accurately leverage up-to-date knowledge over long periods, even in real-world environments where data changes ceaselessly."

The CONDA technology has been applied to AkasicDB, a database product from GraphAI, an AI data infrastructure company founded by Professor Kim, and is set to be commercialized in the fourth quarter of this year.

Lee Da-rae, a graduate of the KAIST School of Computing's master's program, participated in the research as first author, with Professor Kim serving as corresponding author. The findings were presented on September 2 at VLDB 2026, an international conference on databases.