Translation reuse
Preserve useful target text when its source stays the same, and choose regeneration when content or guidance changes require it.
Updating a document or catalog should not require replacing every translation it contains. Translation reuse compares the current source, its previous version, and the existing target so unchanged material can keep its translation.
The format handler must be able to align those units. This is format-aware reuse of supplied content, not a universal fuzzy translation-memory database.
Keep the source and target baselines together #
The current source describes what needs translation now. The previous source identifies what changed. The existing target provides text that can be retained. A target paired with the wrong source revision can lead to unreliable reuse even when both files parse.
For prose, unchanged passages can keep their translations after movement. Unambiguous link changes can be patched while retaining localized targets where appropriate; ambiguous changes need new translation. Structured catalogs have different key and protected-value behavior. Follow the relevant format reference for the material being updated.
If a baseline is invalid or cannot be matched, restore the correct baseline or deliberately regenerate the affected scope. Do not relabel an unrelated source revision merely to make the request valid.
Choose synchronization or regeneration #
Synchronization uses existing target content as the starting point when the source changes. Regeneration requests new output for selected material, including when revised guidance makes the old wording unsuitable.
A localization-engine edit does not rewrite stored translations. If synchronization retains unchanged source units, their targets can still reflect older guidance. To test the new guidance, request fresh translation for a bounded source and locale sample.
A product rename illustrates the difference. Most sentences may remain unchanged while the approved name inside them becomes obsolete. Retaining those targets does not establish that the new naming requirement was applied. Identify the affected content, regenerate deliberately, and compare the result with the new requirement.
Selection and writeback depend on the caller. For the current file workflow, use CLI Push for incremental updates, deliberate regeneration, and baseline handling, or the applicable integration for its selection and delivery rules. Regeneration can consume usage and overwrite accepted text, so limit its scope and inspect the returned changes before accepting them.
Distinguish three forms of reuse #
| Mechanism | What is reused | What still needs examination |
|---|---|---|
| Translation reuse | Existing target text aligned with unchanged source. | Whether changed product facts or guidance invalidate that wording. |
| Context retrieval | Shared terminology, rules, and voice for new generation. | Whether the model selected and followed the intended guidance. |
| Provider prompt caching | A repeated model-input prefix. | New output generation and its remaining costs. |
These mechanisms do not train a model on every correction. A correction can lead to a suggestion for shared guidance, but applying that suggestion is another operation. Inspect the saved change, request a matched translation deliberately, and compare its evidence with the earlier result.
Suggestion generation has its own scoped feedback, batched low-score evidence, debounce, and locale coverage. Generated proposals start pending; applying one saves the guidance change. Read the generation and application policy before treating a review correction as a requirement adopted by every localization engine sharing the resource.