AI research
A research paper titled "AgentOnto: Ontology--Guided LLM Agents for Enterprise Data Analytics" (arXiv:2610.23317), published on October 16, 2026, by researchers from Beihang University, presents a framework designed to resolve semantic misalignment and logical reasoning failures in LLM--driven enterprise data analytics. Enterprise analytics workflows typically require connecting natural--language inquiries across disparate databases, business documentation, and historic task patterns; conventional LLM agents and unstructured retrieval--augmented generation pipelines often struggle with multi--table relationships and complex metric definitions, resulting in invalid query executions and silent semantic errors. AgentOnto addresses this by structuring enterprise knowledge into an operational domain ontology that explicitly maps business metrics, entities, and database constraints. The framework features an "Iterative Ontology Preparation" phase that stress--tests the ontology using auto--generated analytical tasks, alongside an adaptive update protocol with formal correction and retention checks to incorporate user feedback without regressing existing analytical competencies. Evaluated across retail and insurance benchmarks and demonstrated in production settings, AgentOnto significantly improves query accuracy and execution soundness over standard agentic baselines.