AI translation settings
Decide when Glossa translates on its own, which model it uses, and what memory it draws on, project by project.
Each project has its own AI translation settings. Open the project, select Settings, then AI translation. Admins and the account's Owners can change them; everyone else sees them read-only. Changes save as soon as you make them and apply to the next translation run.
Settings marked Available on Team and above are switched off and cannot be turned on while the workspace is on the Free or Starter plan.
Drafting
Auto-translate
"Automatically queue translations when the source text changes."
On by default. With it on, new and changed source strings are translated into every target language as soon as they arrive, whether they come from the CLI, an upload, the API or an integration, and adding a language translates the existing strings into it. With it off, nothing is translated until someone runs it from the editor. See Translate with AI.
A translation you have edited or reviewed is never overwritten by AI, whether a run was started automatically or by hand.
Quality mode
"Use the higher-quality model for every target language in this project."
With Quality mode off, Glossa translates with Glossa Flux, the fast default model. With it on, every target language in the project is translated with Glossa Deep, the quality model. Use it for projects where wording matters more than speed, such as marketing copy or legal text. Team and above.
Quality mode applies to runs started after you turn it on. Translations that already exist are not redone.
Memory
Translation memory
"Reuse an identical machine translation instantly instead of paying to translate it again."
When the exact same source text has been translated before, transglot reuses that translation at no cost instead of translating it again. Exact-match reuse draws on an anonymised, shared pool as well as your own workspace's history. Under the switch, the page shows how many words this project reused this month.
Turn it off and the project neither reuses translations from the pool nor adds its own to it: every string is translated fresh. See Translation memory.
Semantic translation memory
"Find near-matching human-reviewed translations to suggest to reviewers and to ground the AI. Fuzzy matches are never applied automatically."
With it on, transglot also finds previously reviewed translations of similar text and offers them as suggestions. A similar match is only ever a suggestion; it is never written to your project automatically. Reviewers see these matches in the editor under Similar approved translations, and Glossa uses them as reference when it translates. Team and above.
Two settings refine it, and are available only while it is on:
| Setting | What it does |
|---|---|
| Minimum match | "How close a fuzzy match must be before it is worth showing. A legal project may want nothing under 95%; a marketing one is glad of 75%." Exact only shows no similar matches. |
| Match sources | Where similar matches come from: All projects here (every project in this workspace) or This project only. Choose This project only when this project's reviewed translations should not surface in other clients' work, or the other way round. |
What the model can see
Screenshot context
When a key has a screenshot, Glossa is shown it with the string, so it can see where the text appears and write something that fits. It also lets quality checks flag translations that overflow the space they render in. Runs that use screenshots cost more. Team and above. See The screenshot library.
Related settings
These live on the project's other settings pages:
- Route AI translations by confidence, under Settings, then Checks: lets confident translations that pass every automatic check skip review. See Why was my translation flagged for review?
- Voice: the project's brand voice, which Glossa follows. See Brand voice.
- Glossary enforcement, under Checks: how strictly glossary terms are enforced. See The glossary.
If it doesn't work
- A switch is greyed out with Available on Team and above: the account's plan does not include it. See Change your plan.
- Every switch is greyed out: you are not an Admin of this workspace or an Owner of its account.
- Minimum match and Match sources are greyed out: turn on Semantic translation memory first.