Translation pipeline

Choose which transformations and reviews a translation needs, and understand the text each stage can return.

A translation workflow can need more than an initial draft: source cleanup, a more natural target phrasing, review by a person, or shorter wording for an interface. The asynchronous pipeline arranges those operations around a required localization stage.

Each optional stage has a purpose and a cost. Enable it for the work it should accomplish, then inspect its outcome. A job can finish while retaining earlier text after an optional stage fails or is skipped.

If an optional model step cannot complete because of an internal workflow error, the pipeline retains the preceding text and records a stage warning. It does not repeat a potentially completed provider request. Failures to save results or unexpected application errors can still fail the job.

Open the engine's Pipeline tab to enable stages and configure their settings.

Follow the text through the stages #

StagePurposeIf it cannot complete
Pre-editRefine source-language phrasing while asking the model to preserve meaning and structure.Retain the original source.
LocalizationGenerate the target using localization-engine context and the model chain.Fail translation and prevent later refinements.
RephraseRefine target wording before human review.Retain the preceding translation.
Quality gateFind translation errors and use their severity to decide whether human review is needed.Keep review eligible when the assessment is missing or fails.
Human reviewWait for an internal decision or supported professional review.A timeout or unavailable review can retain the preceding translation.
Post-editRefine the human result using translation guidance.Retain human or preceding output.
Content lengthAttempt shorter wording within display-column budgets.Retain prior text, potentially over budget.
Back-translationCompare a reverse translation with the source and attempt correction of consequential drift.Retain the preceding translation on failure.

The web app's creation form defaults to translation only. In the Human review settings, choose a Review level to decide which new translations need a reviewer. You can save the level while human review is off, and turning human review off and back on keeps the saved level. The quality gate runs only when the job's pipeline includes human review.

Choose transformations for a reason #

Pre-edit can help with noisy source content. It is a model cleanup pass, not domain fact-checking, and a completed edit can still be wrong. Carefully edited source may not need it.

Rephrase refines the target using source and localization-engine context before human review. Consider whether another wording change helps the content's purpose. Instructions to preserve markup and placeholders differ from a format handler's protected-key behavior.

Post-edit reconciles human output with terminology, rules, and voice. It requires human review to be enabled and actual human output to exist. The current interface turns post-edit off when review is disabled. Locale bypass, quality-gate bypass, or timeout can leave no human result to refine.

Back-translation translates the latest target in reverse, compares it with the source, and can correct major or critical drift. It adds reverse-translation, comparison, and possibly correction work; it remains a fallible model check, not proof of semantic equivalence.

Optional model passes and professional review can add usage. Compare their results and cost evidence, and check provider entitlements before enabling external review.

Keep review and final acceptance consistent #

Human review is followed by post-edit, content-length processing, and back-translation. Those later stages can change reviewed wording. If publication requires the exact text a person approved, disable later rewriting where appropriate and inspect the final output.

Review can also be bypassed by locale or the quality gate, or expire without human input. The pipeline therefore does not enforce unconditional human approval before delivery. Review policy brings coverage, strictness, provider, timeout, and takeover into one decision. Length limits explains the separate constraints on shortening.

Human review covers the pause, reviewer decisions, quality gate, and timeout. Use Decision pools to create reviewer groups and route the engine's review by language and directory.

Configure jobs through the engine #

Save the stages on the engine before starting a translation job. The job takes its pipeline configuration when it starts, so later edits do not reconfigure work already running. Use a separate engine when two workflows need different policies.

For API configuration, see Pipeline configuration. Sending pipelineConfig replaces the whole engine pipeline configuration; include every stage you intend to enable.

Read what happened #

Use run evidence to inspect stages, warnings, retained text, and skip reasons. Distinguish disabled work from an eligible stage that failed, and a completed transformation from an acceptance decision.

A review can time out while the job continues with earlier text. Establish whether a person supplied a decision and which text survived. Use a defined measurement when a language requirement needs evidence separate from execution status.