Artificial intelligence is becoming increasingly important in businesses, helping employees search through large amounts of documents and answer complex questions.
However, one major problem still limits the use of AI in many industries: hallucinations.
An AI hallucination happens when a system generates information that sounds convincing but is actually wrong.
This can be a serious issue in areas such as law, finance, manufacturing and defense, where inaccurate information can lead to costly mistakes.
Now, researchers at the Korea Advanced Institute of Science and Technology (KAIST) have developed a new database technology that could significantly reduce this problem.
Their system, called AkasicDB, improved AI response accuracy by up to 78% and made some information searches more than 20 times faster than existing methods.
The challenge is that enterprise data comes in many different forms. Some information is stored in documents and reports. Other information is organized in tables, such as sales figures or contract dates. Companies also have data that describes relationships between people, products and organizations.
Most current AI systems struggle because they cannot easily understand all these types of information together. Instead, they often need to search through separate databases and then combine the results. This process can be slow, complicated and sometimes leads to incomplete information being given to the AI, increasing the chances of hallucinations.
To solve this problem, the KAIST team developed AkasicDB, a new database management system that combines three different types of databases into a single platform.
The first is a vector database, which helps AI understand the meaning and context of documents by converting information into mathematical representations.
The second is a graph database, which stores and analyzes relationships among people, companies, products and other entities. The third is a relational database, the traditional table-based system widely used by businesses to organize structured information.
Building on this integrated database, the researchers also created a new retrieval method called Omni RAG. RAG, short for Retrieval-Augmented Generation, allows AI systems to search information sources and use the retrieved information when generating answers.
Conventional RAG systems are effective at finding relevant documents but often struggle when answering more complicated questions that require understanding relationships and specific conditions.
For example, a company might ask an AI system to identify contract clauses related to a particular company that signed agreements last year and explain how those clauses are connected to supply chain issues. Answering such a question requires understanding document meaning, relationships between entities and information stored in tables.
Omni RAG can perform all these tasks simultaneously within a single system. By accessing richer and more complete information, the AI can generate more accurate answers and greatly reduce hallucinations.
In tests, search tasks that previously took more than 21 seconds were completed in less than one second. The researchers believe the technology could become important infrastructure for future AI systems, especially in industries where accuracy and reliability are essential.
As AI agents continue to expand into the workplace, technologies like AkasicDB may help ensure that machines not only respond quickly but also provide answers people can trust.
