pyHermiT

pyHermiT is specified as a Java-free Python and Rust reimplementation of the core HermiT OWL 2 DL reasoner, with a complete pure-Python fallback. It targets Python 3.10+ and uses the Java-free pyowl-core package for shared ontology parsing, immutable views, overlays, composites, and zero-reparse communication with Exact-OM and other consumers.

The 0.2.3 production implementation provides the public reasoner facade, complete pure-Python path, and an optional private Rust backend. It consumes the pyowl-core 0.2 model and can compile its encoded structural-view schema 2 directly in Rust. The normative architecture, compatibility rules, backend completeness requirements, and verification plan begin at specs/README.md. Built runtime artifacts contain no Java, JVM launcher, Java bridge, or reference implementation.

Installation

python -m pip install pyHermiT

A supported native wheel enables Rust acceleration automatically. The universal wheel provides the complete compiler-free Python backend. Neither installation contains, downloads, or starts Java.

Quick start

from pyowl_core import ImportPolicy, LoadOptions, load_snapshot
from pyhermit import Reasoner

ontology = (
    b"Prefix(:=<urn:example#>) Ontology(<urn:example> "
    b"Declaration(Class(:A)) Declaration(Class(:B)) SubClassOf(:A :B))"
)
view = load_snapshot(
    ontology,
    options=LoadOptions(imports=ImportPolicy.RESOLVE_STRICT),
)

with Reasoner(view) as reasoner:
    assert reasoner.ontology is view
    assert reasoner.is_consistent()
    taxonomy = reasoner.class_hierarchy()

See the user guide for backend selection, import resolution, classification and realization queries, timeouts, updates, and error handling. Existing 0.1 deployments should follow the 0.2 migration guide before reusing persisted data or caches.

Release and licensing status

The owner has selected the source-guided implementation mode and LGPL-3.0-or-later, matching the pinned upstream declaration. LICENSE contains the LGPL text, COPYING the GPL text it incorporates, and NOTICE.md the initial upstream attribution. The owner authorized 0.2.3 publication under the existing licensing policy; LIC-001 remains waived without representing that legal review occurred. The authorization is recorded in reports/release/0.2.3-owner-release-override.md. The completed repository audits are under reports/licensing/ and reports/release/artifact-audit.md.

Strict native pipeline

require_native_pipeline_support() checks the native HermiT and pyowl-core receipt capabilities before loading an ontology. Opt in with ReasonerConfig(require_native_pipeline=True) to require native structural validation, metadata, compilation and hierarchy processing. Defaults retain their existing behavior. Strict mode currently supports consistency and class/object/data-property hierarchies on core-validated native snapshot/segment owners. It rejects unsupported owners, Python or verification backends, and Python byte-indexing fallback explicitly. Assertion-local generic queries share the permanent native rules, join plans, role automata, datatype registries and existential templates. Only their facts, clauses, witnesses and mutable tableau state are local. Batches retain at most one active query state; result caches contain bounded Boolean results.

The admitted query syntax includes Boolean class expressions, nominals/equality, known role assertions, existing source literals, universal restrictions and ground existential witnesses. Unknown roles or new literal payloads, newly required inverse-role blocking, nested existential expansions below variable-valued universal restrictions, and global schema/role changes retain the full native rebuild path by default. Strict mode rejects these cases before publication. Realization publication and committed updates are not yet admitted in strict mode. Recurring committed updates have not been established as an Exact workload, so incremental admission remains conditional rather than widening this query envelope.

Reasoner.diagnostics() reports the actual core receipt, metadata validation and domain copies. native_result_validation becomes true after a successful native hierarchy result; native_result_publications counts these publications. A positive capability probe does not replace owner validation or admit an unsupported operation. The native_query_* counters separately record delta loads, local rule plans, peak local records, legacy full-program loads and default full-rebuild fallbacks. Query syntax is limited to 1 MiB per request, 16 MiB and 4,096 items per batch; exhausting a resource bound never enables a fallback.

Documentation

Start with the user guide for backend selection, standalone and shared-view loading, every service family, updates, concurrency, and errors. The API reference enumerates every stable facade member, and the developer guide traces the core-view boundary through the private IR and Python/Rust engines. The documentation index links the normative specifications and machine-readable evidence.

The historical 0.1.1 qualification is recorded in the release report, coverage matrix, and benchmark audit. The owner accepted the remaining licensed W3C, live-reference, and dedicated-performance runs as post-release follow-up for 0.2.0. The release workflow separately requires every configured native target to pass before trusted publication.

Development

Use Python 3.10 or newer and install a compatible pyowl-core:

python -m pip install \
  "pyowl-core>=0.2.1,<0.3" \
  "build>=1.2,<2" "hypothesis>=6.100,<7" "import-linter>=2.1,<3" \
  "mypy>=1.10,<3" "packaging>=24,<27" "pytest>=8.2,<10" \
  "pytest-cov>=5,<8" "ruff==0.15.22" "setuptools==83.0.0" \
  "setuptools-rust==1.13.0" "tomli>=2.0,<3; python_version < '3.11'"
PYHERMIT_BUILD_NATIVE=0 python -m pip install --no-build-isolation -e .
python -m pytest
python -m tools.specs.check_workpackages
python -m tools.specs.check_project
python -m tools.specs.check_links
python -m tools.specs.check_release_gate --require-publishable
ruff format --check .
ruff check .
mypy
lint-imports
cargo deny --manifest-path native/Cargo.toml check --config deny.toml

Build and inspect the compiler-free artifacts with:

SOURCE_DATE_EPOCH=946684800 PYHERMIT_BUILD_NATIVE=0 python -m build
python -m tools.packaging_probe.check_artifact --pure dist/*.whl
python -m tools.packaging_probe.check_artifact dist/*.tar.gz

PYHERMIT_BUILD_NATIVE=auto|0|1 is the only supported build switch:

• auto (default) attempts the optional native extension when Cargo is available and otherwise produces a truthfully tagged Python-only local build;
• 0 never declares the extension and produces the reproducible py3-none-any fallback; and
• 1 requires the locked Rust build and fails instead of silently falling back.

The distribution workflow builds the sdist, universal fallback, and the eight required cp310-abi3 manylinux 2.17, musllinux 1.2, macOS, and Windows x86-64/ARM64 targets. It audits Rust advisories/licenses/sources, ABI, and external libraries; installs on CPython 3.10 and 3.12; compares pure/native metadata and Python payloads, and verifies local-index resolver preference. The target manifest remains configured-awaiting-hosted-validation until those hosted jobs pass. The universal Python fallback is the portable production artifact while additional native wheels complete that hosted validation.