KAIST develops AI search technology that boosts accuracy by 24.5% amid constant data changes

South Korean researchers have developed a technology that preserves AI search pathways even as data is continuously added and deleted, allowing AI systems to accurately retrieve the latest information. The technology is drawing attention for its potential to improve the performance of AI services that handle real-time information changes — including retrieval-augmented generation, or RAG, as well as enterprise knowledge search and product recommendation systems.

KAIST announced Monday that a research team led by Professor Kim Min-su of the School of Computing has developed CONDA — short for Connectivity-Aware Dynamic Index — a dynamic vector search technology that maintains high search accuracy even in environments where data is constantly being added and removed.

Generative AI cannot independently learn information that emerges after its training is complete. To address this limitation, RAG technology has been gaining traction, enabling AI systems to search through recent news articles and internal corporate documents to inform their responses.

Vector search is one of the key factors determining RAG performance. It works by converting the meaning of documents or images into vectors — sets of numbers a computer can process — and then linking semantically similar pieces of information to locate relevant data.

The challenge is that real-world data changes constantly. New news articles, documents and product information arrive while existing data is revised or deleted. As this process repeats, the network of connections between data points gradually breaks down — a phenomenon the researchers call "connectivity collapse."

In simple terms, it is like a destination remaining in place while the roads leading to it disappear one by one. Even if the needed information exists in the database, AI cannot retrieve it once the search path is severed.

CONDA addresses this by considering not only the distance between data points but also whether search paths are properly connected to reach the target information. When new data is added or existing data is deleted, the system preserves critical connections to prevent specific pieces of information from becoming isolated within the search network.

In environments where data is continuously added and deleted, CONDA improved search accuracy by up to 24.5 percent over existing state-of-the-art technologies. Data processing speed was also up to 1.90 times faster.

The research team ran simultaneous search and data update operations for six hours in a large-scale environment containing about 100 million data entries. CONDA maintained the shortest response time and highest search accuracy among all compared technologies, confirming its potential for deployment in real-world large-scale AI services.

The technology's range of applications is broad. It can be applied to RAG-based generative AI systems that must continuously incorporate the latest documents, recommendation systems where user and product information changes in real time, and search functions across news, social media and e-commerce platforms.

Many current search systems handle updated data by storing changes separately or periodically rebuilding the entire search structure from scratch. As data volumes grow, the time and computing resources required for such rebuilds increase significantly. CONDA reduces this burden by updating only the changed portions while preserving core search pathways, eliminating the need for repeated full-scale reconstruction.

The CONDA technology is being applied to AkasicDB, an integrated database product from GraphAI, an AI data infrastructure company founded by Professor Kim. A version of AkasicDB equipped with CONDA-based dynamic vector search is scheduled for release in the fourth quarter of this year.

"The key value of RAG lies in its ability to instantly find and use the latest information that an LLM has not been trained on, exactly when it is needed," Kim said. "We believe this research is significant in that it presents a data infrastructure that allows AI to accurately leverage up-to-date knowledge over extended periods, even as data changes continuously in the real world."

The findings were presented at VLDB 2026, an international academic conference on database research.