A Query.Farm VGI worker for DuckDB.

Grammar, Style & Spell Checking with Auto-Correct in DuckDB

vgi-grammar · a Query.Farm VGI worker · powered by LanguageTool

A VGI worker (Java) that checks grammar, style, and spelling and auto-corrects text as DuckDB SQL functions, powered by LanguageTool. Catalog name: grammar.

INSTALL vgi FROM community; LOAD vgi;
ATTACH 'grammar' (TYPE vgi, LOCATION 'java -jar vgi-grammar-all.jar');

-- One row per issue, with suggested replacements (suggestions is VARCHAR[]):
SELECT rule_id, message, bad_text, suggestions
FROM grammar.grammar_check('She go to school every day.');

-- Auto-correct: apply the first suggestion of each issue.
SELECT grammar.correct('She go to school every day.');   -- 'She goes to school every day.'

-- Count / boolean helpers:
SELECT grammar.grammar_count('She go to school.');        -- 1
SELECT grammar.is_grammatical('This sentence is correct.'); -- true

-- Supported language codes:
SELECT * FROM grammar.grammar_languages();

The worker is a self-contained, shaded fat JAR. Its manifest sets Add-Opens: java.base/java.nio, so a bare java -jar vgi-grammar-all.jar works directly as a VGI LOCATION with no extra JVM flags.

Functions

Function	Kind	Signature	Result
grammar_check	table	grammar_check(text, language := 'en-US')	one row per issue: (rule_id VARCHAR, category VARCHAR, message VARCHAR, "offset" INT, length INT, bad_text VARCHAR, suggestions VARCHAR[])
grammar_count	scalar	grammar_count(text) / grammar_count(text, language)	INT — number of issues
is_grammatical	scalar	is_grammatical(text) / is_grammatical(text, language)	BOOLEAN — zero issues?
correct	scalar	correct(text) / correct(text, language)	VARCHAR — corrected text
grammar_languages	table	grammar_languages()	(code VARCHAR, name VARCHAR) — supported language codes

Behaviour:

• NULL / empty text → grammar_check returns no rows; the scalars return NULL (count/bool) or the input unchanged (correct returns the original, or NULL for NULL input).
• correct applies the first suggestion of each issue non-overlapping, right-to-left (highest offset first) so earlier offsets stay valid; issues without a suggestion are skipped.
• Unknown language → a clear, query-failing error (unknown language 'xx-ZZ'; see grammar_languages() for supported codes).
• The worker catches per-row errors so one bad value never crashes a batch; all logging goes to stderr only (stdout is the Arrow-IPC channel).

Languages

This worker ships the English rule set (org.languagetool:language-en), which provides these language codes:

Code	Name
en	English (generic)
en-US	English (US) — default
en-GB	English (GB)
en-AU	English (Australian)
en-CA	English (Canadian)
en-NZ	English (New Zealand)
en-ZA	English (South African)

grammar_languages() is the authoritative list at runtime. English only is shipped on purpose: each additional org.languagetool:language-* module carries a large rule/resource set and would balloon the ~91 MB fat JAR. To add more languages, declare the desired language-* artifacts in build.gradle.kts and rebuild — the code resolves any language LanguageTool supports with no changes.

Performance / warm-up

A JLanguageTool instance per language is expensive to build (it loads that language's full rule set and tokenizer resources). The worker therefore builds each one once and caches it for the process lifetime, keyed by resolved language code. Because a VGI worker is a long-lived process, this cost is paid once and amortized across every query.

The default language (en-US) is warmed at startup on a background daemon thread, so the worker starts serving immediately and the first default-language query doesn't pay the build cost.

Dependencies & licensing

Component	License	Notes
vgi-grammar (this worker)	MIT	This repository's own code.
LanguageTool (org.languagetool:language-en, languagetool-core)	LGPL-2.1	Grammar / style / spelling engine + English rules. See the LGPL note below.
farm.query:vgi / farm.query:vgirpc	Query Farm Source-Available	The VGI Java SDK (the worker/catalog + RPC API).
Apache Arrow, SLF4J, Log4j-to-SLF4J	Apache-2.0	Arrow IPC transport + logging-to-stderr bridge.

LGPL note for LanguageTool

LanguageTool is licensed under the LGPL-2.1. This worker uses it as an unmodified, standard Maven dependency — it is depended upon and called through its public API, never copied into or patched within this repository. Under the LGPL that is the "using the library" case (not "modifying" it), so vgi-grammar's own code remains MIT and is fine for commercial use.

The standard LGPL obligation is that a recipient of a distributed bundle must be able to relink or replace the LGPL component with a modified version. That is satisfied here because LanguageTool is consumed as an ordinary, swappable Maven Central artifact (org.languagetool:language-en, pinned by version in build.gradle.kts): the build is reproducible from source, the LanguageTool JARs are unmodified upstream releases, and anyone can bump or replace the LanguageTool version and rebuild the fat JAR to relink against a different (e.g. self-modified) LanguageTool. If you ever vendor or patch LanguageTool itself, those changes must in turn be offered under the LGPL.

Building & testing

./gradlew test        # JUnit (in-process table/scalar drivers, real LanguageTool)
make test-sql         # shadowJar + haybarn-unittest over test/sql/*
make test             # both

make test-sql builds build/libs/vgi-grammar-*-all.jar, sets VGI_GRAMMAR_WORKER="java -jar <abs jar>", and runs haybarn-unittest with the worker attached as a VGI LOCATION. Install the runner once:

uv tool install haybarn-unittest
export PATH="$HOME/.local/bin:$PATH"

The VGI SDK and LanguageTool both resolve from Maven Central, so the build is fully self-contained — no mavenLocal, no sibling checkout, no composite build. CI (.github/workflows/test.yml) is a single build-and-test job: JUnit → shadowJar → HTTP boot smoke test → make test-sql.

License

MIT — see LICENSE. LanguageTool is LGPL-2.1 (see the note above).

---

Authorship & License

Written by Query.Farm.

Copyright 2026 Query Farm LLC - https://query.farm