Publications · Preprint, 2026
The Agentic Ontologist: Orchestrating Memory-Augmented Agents for Ontology Engineering
TL;DR Four memory-augmented agents integrate two ontologies: they match classes, extract missing relations, place unmatched classes, and repair conflicting matches. An orchestrator sends each decision to the cheapest tier that can settle it, from label evidence to a local or API model, within a budget.
Highlights
- On the OAEI Mouse–Human anatomy track, matching reaches 0.913 F1 with routing ($0.74 of API calls) against 0.924 with the API model alone ($0.80), 0.839 with evidence alone, and 0.823 with a local Llama 3.1 alone.
- Audits show Llama 3.1 disagrees with the API model too often to be trusted, so the orchestrator stops trusting it: with this local model, routing saves little beyond the free evidence tier, which settles a third of the matching questions.
Overview
The Agentic Ontologist orchestrates four agents over one shared blackboard: Kroma matches classes, an extraction agent proposes missing subClassOf and part_of edges, a reasoning agent places classes that have no match under the other ontology, and a validation agent removes matches that contradict each other. The orchestrator routes each question through evidence, a local model, and an API model in batches, within a budget, audits the local model's confident answers, and recalibrates how far it trusts the local model from where the two agree. Around the agents, it combines RDF/OWL parsing, graph and text embeddings, pluggable retrieval, validation, uncertainty quantification, and Ontology-of-Thought traces, with a tiered query cache that replays repeated and reworded questions.
Cite
@misc{nguyen_2026_agentic_ontologist,
title={The Agentic Ontologist: Orchestrating Memory-Augmented Agents for Ontology Engineering},
author={Lam Nguyen and Ethan Frakes and Hanchao Ma and Ozan Dernek and Roger H. French and Yinghui Wu},
year={2026},
url={https://github.com/lamng3/agentic-ontologist},
}