KAIST Team Debuts AkasicDB to Combat AI Hallucinations and Boost RAG Accuracy

A significant advancement in AI accuracy for enterprise applications has been announced today with the debut of AkasicDB and Omni RAG, developed by a research team at KAIST and the faculty-founded startup Graphy. This new technology directly tackles the persistent problem of AI hallucinations and limited understanding in Retrieval-Augmented Generation (RAG) systems, particularly when dealing with the varied and complex data found in corporate environments. Traditional RAG approaches primarily rely on converting queries and documents into vectors to find semantically similar information. While effective for basic retrieval, this method often falls short when faced with intricate corporate data, which can be dispersed across documents, tables, and relationship graphs. The inability of conventional RAG to comprehensively grasp these diverse data forms leads to a higher propensity for AI hallucinations, where models generate plausible but factually incorrect responses due to insufficient grounding. AkasicDB is engineered to overcome these limitations by integrating the functionalities of vector databases, graph databases, and relational databases into a single, unified management system. This innovative architecture allows users to formulate complex queries that seamlessly combine vector search for semantic similarity, graph traversal for understanding relationships between entities (like people, companies, or products), and relational filtering for structured data. For instance, a user can execute a single SQL/GQL query that leverages all three database paradigms simultaneously. This integrated capability is crucial for enterprise AI agents that need to process real-world corporate data, which rarely exists in a single, homogenous format. By enabling AI to access and interpret semantic information, entity relationships, and tabular data within a unified framework, AkasicDB and Omni RAG significantly enhance the AI's ability to provide accurate and well-grounded answers. This reduces the risk of generating misleading information and improves the overall reliability of AI-powered systems. The research findings were presented as a demo paper at ACM SIGMOD 2026, an international conference in the field of databases, highlighting the novelty and potential impact of this integrated database technology.