Knowledge Graphs & Ontologies

A useful knowledge graph is not a picture of connected circles. It is a managed collection of claims whose identity, meaning, source, and lifecycle are explicit. Begin with a claim such as “Northwind acquired Contoso on 2025-04-03.” The claim contains two identified entities, a typed relation, a date value, and an implicit assertion that someone or some process had grounds to publish it. Graph engineering starts by deciding how each of those parts can be named, compared, queried, challenged, and revised.

RDF provides a deliberately small abstract model. An RDF graph is a set of triples. Each triple has a subject, predicate, and object. Subjects are IRIs or blank nodes; predicates are IRIs; objects are IRIs, blank nodes, or literals. An IRI is not merely a display string. Within the graph it is a global name whose repeated appearance denotes the same resource. A literal carries a lexical form plus a datatype or language tag. The integer literal 7, the string literal “7,” and an IRI ending in /7 are different RDF terms even when a user interface renders them similarly.

The triple model forces a productive separation between identity and labels. The company IRI can remain stable when the company changes its legal name. Human-readable names become claims attached to that IRI, with language, source, and time if necessary. Conversely, identical labels do not establish identity: two people named Jordan Lee require different IRIs until evidence supports a merge. Treating labels as keys is one of the fastest ways to corrupt a graph.

RDF datasets add a default graph and zero or more named graphs. A graph name can identify a source snapshot, tenant, assertion batch, review state, or other application boundary. The RDF specification intentionally does not assign one universal meaning to named graphs, so a production design must document its convention. For example, a dataset may place extracted claims in a graph named for an immutable ingestion batch and place provenance about that batch in a separate metadata graph. A query can then include or exclude sources explicitly.

Blank nodes express the existence of something without giving it a global name. They are convenient for short-lived structures such as an address value assembled inside one document. Their identifiers are local serialization artifacts, not portable identities. Using blank nodes for customers, products, policies, or other records that must survive merging and auditing creates future reconciliation work. Mint stable IRIs for durable domain objects and reserve blank nodes for genuinely local existential structure.

RDF follows an open-world style of representation: a graph normally states what is known, not everything that is allowed. If no triple says that Contoso has a regulator, the graph does not thereby prove that Contoso has no regulator. The absence might mean unknown, not applicable, not ingested, inaccessible, or false. Applications that need required fields and closed lists should add a validation contract rather than pretending graph absence has one fixed interpretation.

Serialization is separate from meaning. Turtle, JSON-LD, N-Triples, and other syntaxes can encode the same RDF graph. Turtle is especially useful for reviews because prefixes shorten IRIs and repeated subjects are compact. JSON-LD helps web APIs participate in linked-data identity, but its context processing must be governed like any other schema. Choose a serialization for the interface while testing the graph it produces, not just the appearance of the document.

Finally, distinguish three layers. The data graph contains domain claims. The vocabulary or ontology defines terms and intended semantics. The operational control plane tracks versions, access, provenance, validation, and releases. Collapsing all three into an unlabeled edge store makes later reasoning unsafe. The graph model is simple; the engineering discipline is deciding what each graph means and who may change it.