Skip to main content

Running translation

The pre-translate dialog​

Start a run from the project or editor with the Translate action. The dialog gives you:

  • AI Model — which provider/model performs the translation.
  • Scope — Empty segments only (the safe default) or Overwrite all.
  • TM match threshold — the minimum fuzzy-match percentage for TM hits to be applied or offered as context.
  • Include TM matches for context — feed fuzzy TM matches to the AI as reference translations.
  • Include TB terms for consistency — inject matched termbase terms into the prompt so approved terminology is used.
  • Include cross-language matches — leverage matches from other target languages in the project.
  • Use project research context — include the project's research notes and client context in the prompt.

Pre-translate dialog

Depending on the project's workflow type, translation may also start automatically when files finish processing.

Choosing a model​

Translation is AI-driven: pick the model per run — OpenAI, Anthropic Claude, Google Gemini, or a local/hosted model in Privacy Mode. Models follow style instructions and use your TM/TB as context. For content that shouldn't touch a model at all, use the TM only (no AI) translation mode.

If a provider fails mid-run, the platform retries and can fall back to an alternative model rather than failing the job.

Privacy Mode​

For content that must not go to the big external AI providers, enable Privacy Mode in Settings → LLM Model Settings, with two Privacy Sources:

  • Platform — hosted TranslateGemma, no setup needed, full GDPR/data residency compliance.
  • Local — your own LLM server: set the server URL and choose the model (TranslateGemma 27B (Specialized), Gemma 3 12B/27B, or any compatible model). Nothing leaves your infrastructure.

While Privacy Mode is active, a banner reminds everyone that translation is running through the private route.

Model settings​

The same settings page also controls:

  • Primary model and fallback models — if the primary provider fails, the run falls back down your list instead of failing.
  • Council Mode — multiple frontier models translate and deliberate together for the highest-stakes content (slower and more expensive; there's a cost warning for a reason).
  • Track Model Usage in History — record which model produced each translation.

LLM model settings

What happens automatically​

Before any AI sees your content:

  • 100% TM matches are pre-applied — segments with exact translation memory matches are filled in and don't consume AI tokens.
  • Repetitions are handled once — repeated segments are translated once and propagated.

During the AI run:

  • Termbase terms matched in each segment are injected into the prompt, so the model uses your approved terminology.
  • Fuzzy TM matches are provided as reference translations.

Translation modes​

When confirming an AI translation run from the editor, you choose a Translation mode:

  • Faithful translation — the default: accurate, close to source.
  • Creative / transcreation — freer adaptation for marketing and brand copy.
  • Dialect / regional variant — target a specific regional variant.
  • Monolingual polish — rewrite the existing text without translating (same-language improvement).
  • Amend existing translation — revise current translations following a revision instruction you supply.
  • TM only (no AI) — apply TM matches at or above the threshold and leave unmatched segments untouched; nothing is sent to any model.

You can also override the model for this file without changing organisation defaults.

Document mode​

For content where segment-by-segment translation loses context (marketing copy, articles, subtitles), toggle Document mode: a whole-document pass for consistency, with TM and TB context inline and rewrite variants. It can also extract candidate terminology for your termbase as it goes.

Prompt rules​

Projects can carry Custom Instructions templates — named, reusable rule sets ("never translate product names", "use formal address") assembled from presets or written freehand. They're injected into every AI run on the project, so per-client conventions don't depend on whoever happens to start the job.

Re-translating​

You can re-run translation on selected segments — for example after changing the termbase, or with custom instructions for a problem batch — without touching segments that are already approved. Choosing Empty segments only in the dialog keeps existing work safe; Overwrite all replaces drafts.

Monitoring progress​

Active jobs appear in a banner with live progress. Note two things that look odd but are correct:

  • The denominator may be lower than the total segment count — segments covered by TM pre-fill and repetitions are excluded from the AI batch count.
  • Large runs are batched by word count, not segment count, so a "batch" can be many small segments or a few huge ones.

Expand the banner's details to see per-stream progress. You'll get a notification (in-app, and by email if enabled) when the run completes.