ทันเอไอ

Researchers develop ISEE system to help AI agents better understand enterprise databases

The Interactive Semantic Enrichment system from Purdue University, Adobe, and UW–Madison enriches data field descriptions with domain knowledge, enabling AI agents to work with databases more accurately

As AI agents powered by LLMs (Large Language Models) are increasingly used to work with data—whether for data understanding, exploration, or querying—a common problem is degraded accuracy, especially with enterprise-grade databases. The cause isn't always the model itself, but rather the field descriptions, which are often vague or incomplete. For example, fields named with company-specific terminology whose true meaning lives in employees' domain knowledge but has never been formally documented.

To bridge this gap, a research team from Purdue University, Adobe, and the University of Wisconsin–Madison, led by Yuan Tian, Yiru Chen, and Rakesh R. Menon and colleagues, has proposed ISEE (Interactive SEmantic Enrichment), published on arXiv under cs.AI/cs.IR. The system works interactively: when given a field description, ISEE evaluates the quality of that description using a scoring mechanism, then gathers knowledge from users and collaborates with them to refine the field's semantic meaning to be more complete.

The research team notes that vague field descriptions directly impact downstream tasks such as entity linking—the task of connecting data to the correct entities—which is exactly the type of work AI agents easily get wrong if they don't understand what each field actually means.

User studies and qualitative evaluation found that ISEE genuinely improves the accuracy of data field descriptions and also reduces users' cognitive load compared to conventional baseline methods—meaning users don't have to exhaust themselves thinking through and documenting everything from scratch.

However, this work is still early-stage research published as a preprint on arXiv and has not yet undergone formal peer review. Detailed experimental figures have not been disclosed in the abstract, so real-world industrial adoption remains to be seen.