Online Ontology Refinement for Analytical Agents via Regret-Bounded Feedback Aggregation in Experimental Serving Environments

Abstract

Large Language Model-based analytical agents deployed in enterprise environments depend on structured semantic layers comprising metric definitions, dimension hierarchies, and entity relationships to ground code generation in domain-specific knowledge and suppress hallucination. These semantic layers are currently maintained as static artifacts by small expert teams, producing two systemic failures: a representation gap, in which the ontology reflects the analytical patterns of a minority of power users while underserving the majority of the user population, and a staleness gap, in which real-world business logic evolves faster than manual ontology updates can accommodate. This article proposes a closed-loop architecture that transforms implicit user feedback signals into automated ontology refinement, golden dataset expansion, and prompt recalibration, with formal convergence guarantees grounded in online convex optimization theory. The framework models the semantic ontology as a structured parameter space and applies regret-bounded gradient updates to both continuous and discrete subspaces, achieving sublinear cumulative regret against the best fixed ontology in hindsight. Evaluation on a simulated multi-tenant analytics platform demonstrates a thirty-one percent reduction in agent error rate, a nine-fold increase in improvement signal capture, and substantially improved representation equity across user segments, with convergence achieved within twenty-one days for the median metric definition.