Quality Check tool

The Quality Check tool compares each block's source with its translation in the language --target-lang names, and reports what it finds. --mode selects how:

rules, the default, runs deterministic checks locally. They cover empty and identical targets, whitespace, doubled words, inline codes, span constraints, characters, length, and, when enabled, consistency across the run.
ai sends each translated block's source and translation to the provider --provider names, which reviews the dimensions in --checks. A hosted provider needs a credential. --provider and --model apply only in this mode.

The tool reads a file that holds both languages, such as XLIFF, or a source file with its translation passed as --target <file>. Findings print on standard output as a table, or as a JSON report with --json. The run writes no file unless -o or --output-dir names one. Each finding is marked as failing or reported, and the run exits 0 whatever it finds. It exits 4 when it compared no translation and reported nothing.

IDqa

SourceBuilt-in

CategoryQuality

Cardinalitybilingual

Requirestarget-language, credentials

Tagsquality, languages

Parameters

Parameter	Type	Default	Description
`absoluteMaxCharLength`	integer	255	Maximum number of characters allowed in any target segment
`apiKey`	string		API key for the AI provider
`charset`	string	ISO-8859-1	Name of the character set encoding to check against (e.g. ISO-8859-1)
`checkAbsoluteMaxCharLength`	boolean	false	Flag target segments that exceed an absolute character count limit
`checkCharset`	boolean	false	Warn if a target character is not included in the specified character set encoding
`checkCodeDifference`	boolean	true	Verify that target segments have the same inline codes as source segments
`checkCorruptedCharacters`	boolean	true	Check for patterns indicating encoding corruption (mojibake
`checkDoubledWord`	boolean	true	Check for consecutive repeated words in target text
`checkDoubleSpaces`	boolean	true	Check for double spaces in target text
`checkEmptySource`	boolean	true	Check for non-empty target when source is empty
`checkEmptyTarget`	boolean	true	Check for empty target when source has content
`checkLeadingWhitespace`	boolean	true	Check for leading whitespace mismatches between source and target
`checkMaxCharLength`	boolean	true	Flag targets longer than a percentage of source character length
`checkMaxWords`	boolean	false	Flag target segments that exceed an absolute limit on the number of words
`checkMinCharLength`	boolean	true	Flag targets shorter than a percentage of source character length
`checkPatterns`	boolean	true	Verify that source patterns have expected corresponding content in the target
`checks`	string[]		Quality checks to perform (e.g. terminology fluency accuracy consistency)
`checkSourceInconsistency`	boolean	false	Flag when different source texts share the same translation
`checkSpanConstraints`	boolean	true	Check non-deletable and non-cloneable span constraint violations
`checkTargetInconsistency`	boolean	false	Flag when the same source text is translated differently across the run
`checkTargetSameAsSource`	boolean	true	Check when target text is identical to source text
`checkTrailingWhitespace`	boolean	true	Check for trailing whitespace mismatches between source and target
`consistencyCaseSensitive`	boolean	true	Whether consistency comparison is case-sensitive
`doubledWordExceptions`	string	sie;vous;nous	Semicolon-separated list of words allowed to repeat (e.g. sie;vous;nous)
`forbiddenChars`	string		Characters that must not appear in target text (e.g. {}[])
`maxCharLengthAbove`	integer	200	Maximum allowed percentage of source length for long text
`maxCharLengthBelow`	integer	350	Maximum allowed percentage of source length for short text
`maxCharLengthBreak`	integer	20	Character count above which text is considered long for the maximum length check
`maxWords`	integer	0	Maximum number of words allowed in any target segment
`minCharLengthAbove`	integer	45	Minimum required percentage of source length for long text
`minCharLengthBelow`	integer	30	Minimum required percentage of source length for short text
`minCharLengthBreak`	integer	20	Character count above which text is considered long for the minimum length check
`mode`	string	rules	How to check quality: deterministic local rules, or an AI provider's review
`model`	string		AI model name
`provider`	string	anthropic	AI provider
`requiredChars`	string		Characters that must appear in target if present in source (e.g. punctuation)
`strictCodeOrder`	boolean	false	Flag differences when codes appear in a different order between source and target
`targetSameAsSourceWithCodes`	boolean	true	Include inline codes when comparing source and target for identity
`targetSameAsSourceWithNumbers`	boolean	true	Include number-only segments in same-as-source comparison

Configure it live

Configuration

Loading form…

Examples

Inline-code and whitespace checks only

Focus on code preservation and whitespace, disabling the length-ratio checks.

checkCodeDifference: true
checkLeadingWhitespace: true
checkTrailingWhitespace: true
checkMaxCharLength: false
checkMinCharLength: false

Enforce an absolute character cap

Report any target longer than 200 characters.

checkAbsoluteMaxCharLength: true
absoluteMaxCharLength: 200

Review with an AI provider

Ask the provider to review fluency and accuracy.

mode: ai
provider: anthropic
checks: [fluency, accuracy]

Processing notes

Operates on translatable blocks; non-translatable structure passes through unchanged.
When a block has no translation in the target language, only the empty-target check applies.
Findings are recorded on the block as a findings annotation, which kapi exec qa prints as its report.
The consistency checks compare each block with the blocks before it in the same run.

Limitations

Length and ratio checks compare character lengths only when source and target are both non-empty.
The regex pattern pairs used by checkPatterns are supplied programmatically, not as schema parameters.
Requires a target language. The tool compares translations in that one language per run.
Run on a translation file alone, the rules mode reads each block as untranslated and reports it as an empty target. Name the source file and pass the translation with --target.
A run with --target pairs one source file with one translation and refuses -o and --output-dir.

← Back to the Tool Reference