rustdl

Sound, performant OWL 2 DL (SROIQ) reasoner in Rust, with Python bindings. No JVM, no subprocess — native classification via PyO3.

rustdl beats HermiT on every measured ORE workload and wins outright against Konclude on Horn-fragment ontologies. See the project README for the full benchmark table.

Install

pip install rustdl

Wheels are published for CPython 3.10+ on Linux (x86_64, aarch64), macOS (Apple Silicon), and Windows (AMD64). Other platforms build from the sdist (needs a Rust toolchain).

Quick start

Prefer a guided walkthrough? See Debugging an ontology with rustdl — an end-to-end QA tutorial (classify → debug() → justify/repair → fix → read inferred facts).

import rustdl

# A small OWL 2 DL ontology ships inside the wheel (gzip-compressed) — no
# download needed. `examples.pizza()` returns its file path (decompressed
# into a per-user cache dir on first use); `examples.PIZZA_NS` is its
# namespace, so class IRIs are PIZZA_NS + local name (e.g. + "Pizza").
from rustdl.examples import pizza, PIZZA_NS, SULO_NS

# Classify. Format is auto-detected from the extension:
# .ofn (OWL Functional), .owx (OWL/XML), .rdf / .owl (RDF/XML), .omn (Manchester).
result = rustdl.classify(pizza())

print(f"{len(result.classes)} classes, {len(result.unsatisfiable)} unsatisfiable, "
      f"complete={result.complete}")
# -> 88 classes, 0 unsatisfiable, complete=True

# Query the computed hierarchy
print(result.is_subclass(PIZZA_NS + "BoxedPizza", PIZZA_NS + "Pizza"))
# -> True
print(len(result.subclasses_of(PIZZA_NS + "FoodMaterial")))
# -> 25

# The pizza ontology is aligned to the SULO upper ontology, so reasoning
# spans both — e.g. a pizza-making timestamp is inferred to be a SULO StartTime:
print(result.is_subclass(PIZZA_NS + "BakingStartTime", SULO_NS + "StartTime"))
# -> True

# Other hierarchy queries (all take full class IRIs):
result.superclasses_of(PIZZA_NS + "Cheese")        # -> list[str]
result.equivalent_classes(PIZZA_NS + "Pizza")      # -> list[str]
result.direct_subsumers(PIZZA_NS + "BoxedPizza")   # -> list[str] (Hasse-direct parents)

Bundled examples

Three real ontologies ship inside the wheel, gzip-compressed (~200 KB total). They classify with no network access — each examples.X() decompresses its ontology into a per-user cache dir ($XDG_CACHE_HOME/rustdl/examples or ~/.cache/rustdl/examples) on first use, then reuses it. Each examples.X_NS is the namespace, so a class IRI is the namespace plus the local name.

import rustdl
from rustdl import examples

r = rustdl.classify(examples.sulo())
print(r.is_subclass(examples.SULO_NS + "StartTime", examples.SULO_NS + "Object"))
# -> True

API

Classification

result = rustdl.classify(path, *, per_pair_timeout_ms=1000, saturation_only=False)
result = rustdl.classify_bytes(data, format="ofn", *, per_pair_timeout_ms=1000, saturation_only=False)

• per_pair_timeout_ms — bound each subsumption test (default 1000; 0 = unbounded). A pair that exceeds the budget is recorded as "not subsumed": sound (never a false subsumption) but the result may be incomplete. When that happens, an IncompleteClassificationWarning is emitted and result.complete is False. Pass 0 for the complete, unbounded classification. The default bounds pathological SROIQ inputs so classification can't hang silently. Conversely, on nominal-heavy ontologies (e.g. the W3C wine ontology) the engines never terminate on the hard pairs and only burn the full budget, so a low value like per_pair_timeout_ms=25 is much faster with no completeness loss (wine: 7.5× faster, identical hierarchy, MISSED=0 vs HermiT).
• saturation_only — skip the tableau entirely; EL-closure-only under-approximation. Dramatically faster on mostly-EL ontologies, and always complete (no tableau ⇒ no timeout).

classify / classify_bytes return a Classification:

One-shot queries

Each parses the file, answers one question, and returns:

rustdl.is_consistent(path)                        # -> bool
rustdl.is_class_satisfiable(path, class_iri)      # -> bool
rustdl.is_subclass_of(path, sub_iri, sup_iri)     # -> bool
rustdl.is_instance_of(path, class_iri, indiv_iri) # -> bool
rustdl.instances_of(path, class_iri)              # -> list[str]
rustdl.realize(path)                              # -> dict[str, list[str]]

realize returns each individual IRI mapped to its most-specific entailed class IRIs.

For repeated queries over the same ontology, prefer classify(path) once and query the returned Classification — each top-level function re-parses.

Inference materialization

rustdl.materialize_inferred_subclass_axioms(path)   # -> list[tuple[str, str]]
rustdl.materialize_inferred_class_assertions(path)  # -> list[tuple[str, str]]

materialize_inferred_subclass_axioms yields (sub, sup) pairs for every entailed subsumption (excluding reflexive, owl:Thing/owl:Nothing, and unsatisfiable classes). materialize_inferred_class_assertions yields (class, individual) pairs. Useful for writing an inferred ontology back to disk.

Errors

rustdl.RustdlError            # base — catches everything from the library
rustdl.ParseError             # the OWL file couldn't be parsed
rustdl.UnsupportedAxiomError  # HasKey, role chains > length 2, etc.
rustdl.UnknownClassError      # an IRI argument isn't a declared class

try:
    result = rustdl.classify("ontology.ofn")
except rustdl.ParseError as e:
    print(f"bad input: {e}")
except rustdl.RustdlError as e:
    print(f"reasoning failed: {e}")

Soundness & coverage

rustdl is sound: every reported subsumption is a genuine entailment (FP=0 against Konclude on the validation corpus). Completeness is partial — the default classifier is empirically near-complete across the measured corpus but not provably complete on all of SROIQ. saturation_only and per_pair_timeout_ms are sound-but-incomplete by construction.

Data-property and datatype axioms outside the recognized preprocessing patterns are silently dropped (a sound under-approximation). HasKey and role chains longer than length 2 raise UnsupportedAxiomError. SWRL rules are skipped.

See the project documentation for the full coverage matrix, soundness contract, and architecture notes.

License

Apache-2.0 OR MIT.