The integration pattern “knowledge-graph retrieval as grounding” combines symbolic and neural AI in the following way: a knowledge graph supplies the factual foundation over which a large language model retrieves and reasons in multi-turn reasoning steps. The language model carries the natural-language understanding and generation; the knowledge graph carries the structured, auditable factual situation.
Representatives include, among others, Graph-R1 (2025) with lightweight knowledge-hypergraph construction and retrieval as a multi-turn agent-environment interaction; ToG-2 (Think-on-Graph 2), which tightly couples context retrieval and graph retrieval; and graph-constrained reasoning, which embeds the KG structure into the decoding process of the language model.
In terms of personhood ontology, this pattern is a special case of neuro-symbolic AI. It is related to Retrieval-Augmented Generation (RAG), but more structured: instead of a flat document collection, a curated graph with explicit entities and relations serves as the source.
Ontological classification: Superordinate concept: neuro-symbolic AI.
Methodological point
The pattern addresses the question of hallucination in large language models from a different direction than “ontology as constraint.” Instead of structurally restricting the model during training, one gives it, at answer time, an explicit factual structure. The model is then no longer dependent on the distribution in its training data, but on the structure of an auditable graph.
The price here too is dependence on the quality of the graph: what is not in the graph, the system cannot retrieve; what stands wrongly in the graph, the system disseminates with the authority of a source. The maintenance of the knowledge graph thereby becomes a critical component — as already in the classical expert systems of the 1980s, only now in front of an incomparably more demanding inference layer.
Connecting point to the present ontology
The pattern has a concrete resonance with the personhood ontology here: a neuro-symbolic system could use the personhood ontology as a knowledge graph in order to answer, for instance, questions in the history of philosophy, in bioethics, or in personal ontology in a more structured manner. It would thereby not understand the personal-ontological distinctions — understanding is reserved to personhood — but it could reference them consistently and secure them against its own generative tendency toward smoothing-over.
Sources: Generated by querying the Personhood ontology.
Further sources:
- Graph-R1 et al. (2025): several arXiv publications on KG-supported LLM retrieval.
- Lewis, Patrick et al. (2020): Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. Advances in Neural Information Processing Systems 33.
- Pan, Shirui, Luo, Linhao, Wang, Yufei, Chen, Chen, Wang, Jiapu & Wu, Xindong (2024): Unifying Large Language Models and Knowledge Graphs. A Roadmap. IEEE Transactions on Knowledge and Data Engineering Vol. 36, Issue 7, pp. 3580–3599. DOI 10.1109/TKDE.2024.3352100.