pylat_ru

pylat_ru is a native Python implementation of the Russian checking pipeline and rule engine used by LanguageTool. Production use requires no Java/JRE, LanguageTool server, network service, Natasha, or pymorphy.

The ordinary (non-language-model) Russian pipeline has full observable parity on the committed upstream and differential evidence suites for LanguageTool 6.8, pinned at e807fcde6a6506191e1470744d2345da28c26be6. This is deliberately narrower than a claim of compatibility with every LanguageTool feature: RussianConfusionProbabilityRule remains LANGUAGE_MODEL_DEFERRED because it depends on LanguageTool's language-model subsystem.

pylat_ru is an independent project and is not affiliated with or endorsed by LanguageTool.

Installation

Install the first public alpha from PyPI:

python -m pip install pylat-ru==0.1.0a0

To opt into the newest prerelease, use python -m pip install --pre pylat-ru. For source development, use python -m pip install -e ".[dev,release]". The only production dependency is regex; build, test, and release tools are optional.

Basic usage

from pylat_ru import LanguageToolRU

tool = LanguageToolRU()
for match in tool.check("Ученик решил задать тест учителю."):
    print(match.rule_id, match.message)
    print(match.offset, match.length, list(match.replacements))

The stable primary API is LanguageToolRU, RuleMatch, and __version__. LEVEL_DEFAULT and LEVEL_PICKY are stable convenience constants. Existing exported tokenization, analysis, tagging, disambiguation, and synthesis classes remain available as an advanced/provisional surface; they do not yet carry the same stability promise.

Checking levels and rule control

check() uses DEFAULT unless asked for PICKY. Unknown levels fail explicitly.

from pylat_ru import LanguageToolRU, LEVEL_DEFAULT, LEVEL_PICKY

tool = LanguageToolRU(rule_config={"TOO_LONG_SENTENCE": {"maxWords": 4}})
text = "Один два три четыре пять."
assert not tool.check(text, level=LEVEL_DEFAULT)
assert any(m.rule_id == "TOO_LONG_SENTENCE" for m in tool.check(text, level=LEVEL_PICKY))

Rules can be enabled or disabled by their LanguageTool ID:

from pylat_ru import LanguageToolRU

enabled = LanguageToolRU(enabled_rules=["MORFOLOGIK_RULE_RU_RU_YO"])
assert any(m.rule_id == "MORFOLOGIK_RULE_RU_RU_YO" for m in enabled.check("Ежик и елка."))

disabled = LanguageToolRU(disabled_rules=["zadat_test"])
assert not any(m.rule_id == "zadat_test" for m in disabled.check("Ученик решил задать тест учителю."))

rule_config is validated and unknown rule IDs or option names fail explicitly. A supported spelling example is {"MORFOLOGIK_RULE_RU_RU": {"conf_ru_Value": 1}}; a supported length example is shown above.

Match fields and offsets

The primary RuleMatch fields are rule_id, category_id, message, offset, length, and replacements. Suggestions are ordered and both order and duplicates are meaningful. rule_id is the base upstream ID; full_rule_id identifies the physical variant where upstream exposes one. Additional metadata fields are retained in the public API snapshot.

offset and length are Python Unicode code-point indices into the original input and can be used for normal Python slicing. utf16_offset and utf16_length are UTF-16 code-unit coordinates for interoperability with Java LanguageTool and UTF-16 clients.

Errors, compatibility, and updating upstream

Missing or corrupt packaged resources fail during initialization/use; there is no network or repository fallback for an installed wheel. Unknown checking levels and configuration keys raise explicit exceptions. Compatibility evidence lives under compat/, and the controlled pin update process is documented in docs/upstream_update.md.

License and provenance

Project code is licensed under LGPL-2.1-or-later. Shipped LanguageTool-derived rules, dictionaries, and other resources retain their upstream attribution. See LICENSES.md and license_inventory.json for the recorded evidence; no unresolved BLOCKED_LICENSE_REVIEW resource is intentionally shipped.