The Design¶

Ontoplexis is built on one idea: OWL/XML is the OWL 2 structural specification already serialized as a tree, so a document-to-graph mapping needs no vocabulary — just a generic walker.

The problem it avoids¶

Making OWL queryable usually means choosing a model and writing per-construct mapping code: one class (or table, or handler) for each of OWL 2's ~70 axiom and expression types, on every side of every boundary. That code is the system: it must be maintained in lockstep with itself, and every round-trip guarantee is a promise kept by hand.

The other common shortcut — loading the ontology's RDF triples directly — gets a graph cheaply but the wrong graph: OWL's RDF encoding reifies annotated axioms, chains collections through rdf:first/rdf:rest, and scatters restrictions across blank nodes. The query semantics you want are the structural specification, not its RDF encoding.

The mechanical mapping¶

In OWL/XML, the element name is the construct kind, attributes are its scalar values, and child order is its argument order. So the mapping is:

Relative IRIs are resolved against xml:base at parse time, so the stored graph is always absolute and base-free.

The only OWL-specific knowledge in the system lives in two small data tables: one that decorates edges with query-facing roles (sub, super, property, filler, operand, annotation, …) so queries read naturally, and one that derives the one-hop convenience edges described below. Roles are decoration; position is what round-trips. See Edge Roles.

The derived convenience layer¶

The structural graph is faithful but verbose: every binary fact — "Dog is a subclass of Animal", "rex has type Dog", "Dog's label is Dog" — is a two-hop join through an axiom node. A third table, D, materializes exactly those binary relations as one-hop edges (see Derived Relations).

D is a cache over asserted structure, not inference: each derived edge is a mechanical collapse of a construct whose OWL 2 → RDF mapping is a single triple. graph() and serialization never read it, so round-trip fidelity is untouched by construction. Every projection built by build_projection or save_projection derives it automatically; after authoring or migrating a projection with raw Cypher, call derive_edges to refresh it.

Where OWLAPI fits¶

Only in the test suite. The hard, format-specific work — parsing Turtle or RDF/XML and executing the RDF-to-structural reverse mapping — belongs to OWLAPI, the reference implementation, and the tools built on it (ROBOT, Protégé). Ontoplexis does not wrap them: documents in other serializations are pre-converted to OWL/XML with whichever of those tools you already use, and reasoning (materializing inferred axioms) happens the same way, before loading. The package itself is pure Python.

The test suite downloads ROBOT, the OBO community's OWLAPI CLI, as a development-only test oracle. It never ships, and the repository contains no Java code at all.

Round trips are proven, not promised¶

parse_owlxml(serialize_owlxml(parse_owlxml(x))) reproduces the same graph by construction, and the test suite goes further: it converts documents to canonical functional syntax through the OWLAPI oracle before and after both the document path and the database path and requires byte equality. If the walker ever mishandles a construct, the reference implementation fails the build.

Non-goals¶

• No typed axiom classes in Python. The structural spec lives in the graph shape; typed models live in the sister package ontophora.
• No runtime format conversion, reasoning, or profile checking. OWLAPI is the test oracle, not a dependency.
• No RDF-encoded graph. The derived table D gives the one-hop binary relations RDF users reach for, but as a cache over the structural graph — never reified axioms, rdf:first/rdf:rest chains, or blank-node scatter.
• No materialized transitive closure. Deliberate: the closure can be quadratic in hierarchy size and doubles the cache-invalidation surface, while bounded recursive Cypher over D already answers reachability queries. Traverse instead of materializing.