[jdsika]

Mirror of linkml#4156, same head. Review there; this PR tracks the fork stack.

One rule decides whether a value of a type becomes an IRI, a blank node or a literal in RDF. linkml_runtime.utils.rdf_terms.type_term is shared by gen-jsonld-context, gen-owl, gen-shacl, gen-shex, gen-yarrrml and the RDF dumper and loader.

Example

types:
  FilePath:
    typeof: string
    uri: xsd:anyURI
slots:
  homepage: {range: uri}        # a link
  file_path: {range: FilePath}  # a URI reference held as text, may be relative

For homepage: https://alice.example.org/ and file_path: ./docs/cv.pdf:

	main	this PR
gen-jsonld-context	homepage → "…"^^xsd:anyURI literal, file_path → same literal	homepage → @id, file_path → xsd:anyURI literal
RDF dumper	homepage → IRI, file_path → <./docs/cv.pdf> (not an IRI, RDF 1.1 §3.2)	homepage → IRI, file_path → "./docs/cv.pdf"^^xsd:anyURI
gen-shacl	both sh:nodeKind sh:IRI	sh:IRI / sh:datatype xsd:anyURI
gen-owl	both owl:DatatypeProperty	owl:ObjectProperty / owl:DatatypeProperty
RDF loader (strict)	rejects the dumper's own IRI for homepage	round trip

So on main, data written through the generated context fails the generated shapes, and the dumper writes relative IRIs. The --xsd-anyuri-as-iri option only moved the problem: with it, ./docs/cv.pdf is resolved against the reader's location.

The rule

Type	RDF term
uri, uriorcurie and types derived from them	IRI (CURIEs expanded)
objectidentifier (shex:iri)	IRI
nodeidentifier (shex:nonLiteral), types with uri: rdfs:Resource	IRI or blank node
every other type, e.g. typeof: string with uri: xsd:anyURI	literal of its datatype, as written

The rule reads the inherited, expanded datatype, so its spelling does not matter, and it follows the typeof chain, as gen-python does for the Python base class (ImageLink(URI) vs FilePath(str)). uri and uriorcurie are recognised by name and definition, not by from_schema, which merging imports rewrites. An explicit datatype on a derived type wins (typeof: uri with uri: xsd:string is a string literal).

Standards

• RDF 1.1 Concepts §3.1: IRIs, literals and blank nodes are distinct; §3.2: "IRIs in the RDF abstract syntax MUST be absolute".
• XSD 1.1 §3.3.17: "An anyURI value can be absolute or relative". A URI reference held as text is an xsd:anyURI literal, not an IRI.
• OWL 2 §5.8.1: a property is an object or a data property, so a slot needs one term kind, not one per value.
• SHACL §4.1.3 node kinds; §4.8.3 sh:in compares terms exactly; ShEx 2.1 §5.4.2 iri, bnode, nonliteral.

LinkML contract and history

• The spec (06mapping) maps a type value to Literal(V, Type) and coerces context terms to @type=<s.type.uri>; types.yaml says a uri value "is treated as a literal of type xsd:anyURI unless it is an identifier". cmungall chose literals in Clarify behavior of uriorcurie in rdf dumpers linkml/linkml#450, feat(shacl): improve type generation linkml/linkml#1300 and Fix xsd:anyURI range in JSON-LD generator linkml/linkml#2242.
• In the linkml/linkml#3215 review (2026-02) he asked for "curies and uris … as RDF nodes rather than literals"; Emit URIRef for xsd:anyURI values in RDF serialization linkml/linkml#3221 did that for the dumper and gen-shacl only, keyed on the datatype xsd:anyURI.
• feat(generators): add --xsd-anyuri-as-iri flag for cross-generator IRI consistency linkml/linkml#3292 added --xsd-anyuri-as-iri as an interim option for the context and OWL. Make it possible to map a URI to an RDF Node linkml/linkml#2495 and Using wrong XSD type for URIorCURIE linkml/linkml#2213 are open.
• This PR applies the 2026 direction in every generator, keeps the literal reading for types not derived from uri/uriorcurie, and asks for that decision. The spec and types.yaml text then need a linkml-model update.

Behaviour changes

• gen-jsonld-context, gen-owl: uri/uriorcurie slots and derived types become @id and owl:ObjectProperty by default (before: with the option; OWL also had owl:allValuesFrom xsd:string). The metamodel and Biolink context snapshots change accordingly; OWL type_objects mode is unchanged.
• gen-shacl, gen-shex, dumper: a string-derived xsd:anyURI type is a literal (before: IRI).
• Node types: objectidentifier is sh:IRI (was sh:BlankNodeOrIRI); nodeidentifier and rdfs:Resource types are IRIs or blank nodes in the context, dumper, OWL, ShEx and YARRRML (before: ^^shex:nonLiteral literals that failed SHACL and reloaded as None). Closes nodeidentifier values are written as shex:nonLiteral literals by the JSON-LD context and rdflib_dumper, so they never satisfy gen-shacl #51.
• gen-shacl: sh:in values and sh:defaultValue take the slot's term kind; a curie slot with subproperty_of was unsatisfiable.
• Dumper: a typeof: uri type without its own uri was a plain string literal; now the inherited datatype applies. Relative values of IRI types raise ValueError.
• Loader: accepts IRIs for node-kind ranges in strict mode (cast_literals=False) and returns _:label for blank nodes in IRI-or-blank-node slots.
• Fail fast: a type datatype with an undeclared prefix is an error (xsd, shex, rdfs are known).
• --xsd-anyuri-as-iri: deprecated in 1.13.0, no effect, removal scheduled for 1.14.0. To keep a uri slot as text, give it a string-derived xsd:anyURI type. SEMICeu/core_person_linkml's three range: uri slots would need this.

Alternatives not chosen

• Always literal (spec reading): reverses Emit URIRef for xsd:anyURI values in RDF serialization linkml/linkml#3221.
• Always IRI (datatype xsd:anyURI, as Emit URIRef for xsd:anyURI values in RDF serialization linkml/linkml#3221): relative references, which are valid xsd:anyURI, cannot be IRIs.
• Decide by value: OWL 2 DL needs one kind per property.
• Keep the option or add a per-slot one: the type already states it; the option was interim.
• New iri type: a metamodel change; not favoured in the feat(generators): add --xsd-anyuri-as-iri flag for cross-generator IRI consistency linkml/linkml#3292 review.

Not in this PR

• curie stays an unexpanded xsd:string literal although types.yaml says it MUST be expanded in RDF; JSON-LD cannot expand inside a literal, so it needs a decision (Using wrong XSD type for URIorCURIE linkml/linkml#2213).
• Validation: JSON Schema has format: uri only when the induced base is URI; the runtime URI class accepts relative references; the loader compacts IRIs to CURIEs also for uri slots.
• Runtime Identifier.is_valid rejects blank-node labels such as _:b1, so nodeidentifier blank nodes work through the context and the loader only with rdfs:Resource types until that is fixed.
• linkml-model: metamodel context regeneration, spec and types.yaml text.
• Release: linkml needs linkml-runtime of the same release for the new module (floor bump at release, as for earlier runtime additions).

Testing

• test_uri_term_kinds.py (61): context, SHACL, OWL, ShEx and YARRRML output; dumper and context data paths validated with pyshacl; loader round trip and blank nodes; wrong-kind negatives, relative links, curie with subproperty_of. 31 fail on main.
• test_rdf_terms.py (35): datatypes, spelling, identity in own and merged schemas, prefixes, values.