RoboData: Toward Trustable Question Answering over Ontologies through Metacognitive Agentic Epistemology

Large Language Models (LLMs) have revolutionized Natural Language Processing (NLP), simplifying knowledge extraction from structured and unstructured sources. Despite pervasive usage in various applications, from Question Answering (QA) to goal-driven reasoning, they tend to produce hallucinations, factually incorrect responses that hindering accuracy and explainability of Knowledge-Based QA (KBQA) tasks. To address this limitation, we introduce RoboData, an Agentic AI approach to verifiable knowledge extraction and reasoning over structured ontologies like Wikidata. Through metacognitive self-reflection and goal-directed commonsense reasoning, a LLM-based epistemic agent dynamically self-orchestrates a query answering process. Unlike traditional information retrieval systems, the proposed architecture incrementally builds a local knowledge graph from remote knowledge sources to answer a natural language query with traceable facts, highlighting a ”support set” for each claim, a set of nodes and edges in the local knowledge graph that backs the generated claims. The resulting accumulated knowledge provides an intermediate explainability layer, providing a reliable epistemic substrate for using trusted ontologies in goal-driven query answering applications, such as robotic planning and semantic map enrichment. Code is available at: https://emanuelemusumeci.github.io/RoboData/