Sources

The Gruber 1993 paper is the source of the working definition of an ontology that the field still uses: an explicit specification of a conceptualization. The phrase has two load-bearing parts. "Explicit specification" means the schema is written down in a form a machine can read, not held informally in the heads of domain experts. "Conceptualization" means the schema captures an abstract model of the domain, not the surface text of any particular document.

The translation-approach framing in the paper is also worth knowing. Gruber treats an ontology as a contract between knowledge sources written in different formalisms, with the ontology serving as the shared vocabulary that lets them exchange data. This is exactly the role an ontology plays in a modern GraphRAG pipeline: it is the contract between the LLM extractor, the storage layer, and the query layer.

RDF 1.1 is the W3C Recommendation that defines the triple as the atomic unit of knowledge representation on the semantic web. Everything in an RDF graph is a triple of (subject, predicate, object), and every other semantic-web standard sits on top of this substrate. The abstract syntax in the specification is what all serialization formats (Turtle, N-Triples, RDF/XML, JSON-LD) materialize.

The article's claim that "the discipline is in the schema, not in the serialization format" reflects a practical reading of RDF: production GraphRAG systems rarely exchange data in formal RDF, but the conceptual model of typed triples flows from this specification.

OWL 2 is the W3C Recommendation that defines the expressive constraints layered on top of RDF: cardinality (a Well has exactly one Operator), transitivity (if A is part of B and B is part of C, then A is part of C), equivalence and disjointness, inverse properties, and class expressions that go beyond simple subclassing. The Direct Semantics is what description-logic reasoners implement, and it is what enables the inference and consistency-checking queries the article describes.

The OWL 2 profiles (EL, QL, RL) trade expressiveness for computational tractability. Most production ontologies pick the smallest profile that supports the constraints they actually need, because reasoning over a full-OWL ontology is decidable but expensive. The constraint examples in the article (a Well has exactly one current Operator, a Formation is located in exactly one Basin) are within the OWL 2 RL profile, which is the one most graph databases support natively.

SKOS is the W3C Recommendation designed for thesauri, glossaries, and controlled vocabularies rather than for logical reasoning. The constructs (broader, narrower, related, prefLabel, altLabel) capture vocabulary structure without committing to OWL-grade semantics, which makes SKOS the right tool when the ontology's job is to name things consistently rather than to reason over them.

The synonym-table discipline the article describes (ExxonMobil, Exxon Mobil Corporation, XOM, and Exxon all resolving to the same Operator instance) maps directly onto SKOS prefLabel and altLabel. Most production ontologies use SKOS for the vocabulary layer and OWL for the schema layer, with RDF as the common substrate.

Guha, Brickley, and Macbeth's Communications of the ACM article is the reference history for the schema.org effort. The paper documents the pragmatic principles that drove the design: web-scale adoption was prioritized over depth of formal modeling, the type vocabulary was kept deliberately shallow to encourage uptake by non-specialists, and search-engine markup (Google, Microsoft, Yahoo) was the primary consumer.

The relevance to GraphRAG is the trade-off schema.org documents. A general-purpose ontology with shallow types reaches wide adoption but loses domain precision. A domain ontology with deep types serves a narrow community well but does not transfer. The article's claim that schema.org is a useful reference but rarely sufficient alone for domain work reflects this trade-off, and the same logic explains why most production GraphRAG systems do not use formal RDF or SPARQL: the discipline is in the schema design, not in the serialization format.

The Edge et al. Microsoft GraphRAG paper is the most-cited reference implementation of the LLM-extracted-graph pattern. The paper's indexing pipeline runs an LLM per chunk to extract entities and relationships, then merges the extracted fragments into a corpus-level graph and runs hierarchical Leiden community detection to partition it.

The relevance to the article's argument is that Microsoft's reference implementation pairs an LLM extractor with prompt-level type constraints. The extraction prompt enumerates the canonical entity types and relationship types the model should use, which is exactly the role an ontology plays in a typed extraction pipeline. The graph's quality is bounded by the prompt's quality, which is the operational fact the article's "the ontology is what makes the LLM's extractions reproducible" framing rests on.

HippoRAG shares the upstream extraction pattern with Microsoft GraphRAG (LLM-extracted typed triples) but replaces the downstream community-summarization step with Personalized PageRank over the graph. The paper reports gains of up to 20 percent on multi-hop question answering and 10 to 30 times cheaper retrieval than iterative methods.

The relevance to the article is structural: HippoRAG demonstrates that an ontology-grounded extraction can serve multiple retrieval primitives. The same typed graph supports community-summary aggregation (Microsoft), Personalized PageRank ranking (HippoRAG), and dual-level retrieval (LightRAG). The ontology is the shared substrate; the retrieval primitive is a separate design choice.