A new arXiv paper introduces ISEE, short for Interactive Semantic Enrichment, for improving the descriptions of database fields used by LLM-based data agents. The system targets a common problem: many fields are ambiguous because important domain context lives with users, not in documentation.
ISEE measures the quality of a field description, gathers domain knowledge, and collaborates with users to enrich the semantics. Better descriptions can then support downstream tasks such as entity linking, retrieval, and data exploration.
The authors evaluated the system through a user study, automated user simulation, quantitative tests, and a case study. They report lower cognitive load, higher description quality, and improved downstream task performance.
The work highlights a practical limit of data agents. Model capability is not enough if the data schema is unclear; useful automation may require tools that draw missing meaning out of human domain experts.