Localization Engine Suggestions are concrete edits to a localization engine – a glossary entry, an instruction, or a brand-voice change – proposed by the platform from its own AI Reviewer verdicts. When reviews on a translation come back low, the platform suggests a specific edit to the localization engine's glossary, instructions, or brand voice, with the reasoning attached. Apply writes the edit into the localization engine and the next translation runs with the fix in place. Dismiss drops the suggestion.
An auto-suggestions toggle on the localization engine's Reviews tab runs the analysis in the background – a run of low scores quietly produces suggestions waiting in the Suggestions tab, plus a notification.
Suggestions are addressable from code through the Localization Engine Suggestions API. Free-text feedback – "the German copy sounds too stiff", "stop translating the product name" – goes in the same way and returns the same kind of structured suggestion, with list, apply, and dismiss endpoints alongside.
Prompt caching: repeat translations get cheaper and faster#
The stable parts of every translation prompt – the system prompt, the localization engine's instructions, the brand voice – don't change between requests for the same localization engine and locale. They now reuse the LLM provider's prompt cache as cache tokens, billed at a fraction of the regular input-token rate – the discount flows through to the per-request cost on each call. The per-request glossary and the input text are still sent fresh. Cache reads and writes show up as their own fields in each translation's usage breakdown and as their own series in the token consumption chart.
Also shipped#
- Content length enforcement – a new optional pipeline step that rewrites translations to fit a column budget – a fixed cap, the source's width, or a tolerance band around it – so translated copy doesn't overflow notifications, buttons, or tight layouts.
contextfield on localization requests – an optional payload-level hint (product surface, audience, purpose) that the LLM reads alongside the input. Complements the per-keyhintsfield for one-off guidance that isn't part of the localization engine's configuration.- Notification preferences split into independent in-app and email toggles per notification type.
