How it works

Follow a content change from its source through saved translation guidance to a result you can use and check.

When you write and revise content, you decide what you want readers to understand, what deserves more explanation, and which arguments would persuade them. Translation should convey that meaning and persuasive intent just as precisely in another language. Lingo.dev uses language models to translate your content, with saved guidance about your product, terminology, and voice.

You supply the source and the languages you need, and select the localization engine that should translate them. An API caller identifies it with engineId; a CLI or integration uses the localization engine selected in its configuration. That choice determines the models and shared guidance available for the request. The translated content returns through the tool that requested it, with quality measurements and human review following the paths you configure.

Start with the content and its destination language #

The content you provide is the source. Each translated version is a target. Source content can come from a file, structured data in an application, or a connected content system. The requested source and target locales identify the languages involved.

A locale identifies the language and, when needed, its region or script. Choose the intended audience explicitly: a request for pt-BR names Brazilian Portuguese rather than leaving the region unspecified. The source and target locales also help Lingo.dev select the applicable translation guidance. Locale resolution explains those matching rules.

The content's structure matters alongside its language. An application uses message keys to find text and variables to insert values. A document contains headings, links, and formatting. An API caller supplies structured data directly; the CLI and format handlers extract translatable text from files and reconstruct their targets. Each path needs to return content that the receiving system can still use.

Keep the source revision with the output when storing translations or investigating them later. Once an author changes the meaning, a translation that was correct for the previous version may no longer express the intended message.

Save the knowledge that later requests will need #

A language model knows how to work with language. It does not automatically know your product's terminology, the requirements your team has agreed on, or how your company speaks to readers. A localization engine keeps model choices and links to reusable guidance, so each request can use that knowledge without sending it all again.

Different kinds of guidance answer different questions:

  • A glossary records terms, their intended meaning, and preferred wording or names to leave untranslated. It helps the model distinguish a product concept from another use of the same word.
  • A ruleset contains translation instructions. It can express recurring requirements that a term list cannot, including distinctions in meaning that the translated content must preserve.
  • A brand voice describes how the company speaks to its readers, including tone and formality.

Glossaries, rulesets, and brand voices belong to your organization and can be attached to multiple localization engines. Related workflows can share terminology while keeping different model or voice choices where they need them. Editing a shared resource can affect every localization engine that uses it, so inspect its attachments when changing guidance.

For each request, Lingo.dev selects relevant glossary terms, matching instructions, a voice variation, and model candidates from the chosen localization engine. A JSON document can also supply format.options.context describing the audience or purpose, and format.options.hints for individual keys. This lets you keep shared knowledge in the localization engine while explaining what is particular to the content being translated. Translation context covers those choices in detail.

The creation form's defaults #

The web app's creation form starts with the following defaults. You can change the models and select existing guidance before completing creation.

SettingForm default
ModelsA recommended primary model and fallback cover all locale pairs. You can change the models or configure a particular pair separately.
Glossaries, rulesets, and brand voiceNo attachments selected by default. Select existing resources during creation or attach them later.
PipelineTranslation only. Optional rewriting and human-review stages start disabled.
Glossary and instruction scoring; suggestionsAutomatic scoring, suggestion generation, and automatic application of suggestions start disabled.

Existing localization engines may have different saved settings. Inspect the one your caller will use rather than assuming its configuration still matches the defaults. Create and configure a localization engine covers model choices and attachments.

Generate the target and return it to the caller #

The selected language model receives the source and applicable context. If the first configured model cannot complete the request, the localization engine can try its fallback. A fallback supplies another attempt; it does not compare both translations and choose whichever has the highest quality score.

With the API, you create and start a translation job, then download its completed targets. JSON output keeps the document's message keys, and your application decides where to store or use it. With the CLI, push and pull perform this file workflow for you. An integration has its own writeback behavior, such as opening a translation pull request. Integrations explains what each connection reads and writes.

The returned structure connects translation to the rest of your work. Check that the application can load the catalog, substitute its variables, or render the document's links and formatting. These checks establish that the output is usable by the receiving system; measuring its language quality requires a different kind of evidence.

Check the result in the workflow that uses it #

Choose measurements around what matters in your content. Checking preserved meaning requires the source and its translation, with enough context to understand what the author intended. A spelling check can work from the target alone. The evidence differs because the questions differ. The result is useful when you can see what was checked and inspect the passage behind the answer.

The engine's pipeline can include human review before the job produces its final output. Read review policy for coverage, routing, and timeout behavior: a completed target can have bypassed review or continued after a timeout. Independent quality measurements run separately, so translation completion is not a completed quality report or proof of a person's approval.

Measurements can also examine translations produced outside Lingo.dev. You can keep an existing generation system and use Measure to find the supported way to inspect its output. The measurement answers a defined question about the supplied text; your team decides what the result means for its use.

If a review reveals a recurring terminology problem, correct the current translation and consider a separate glossary or instruction change for future work. Updating shared guidance does not rewrite every earlier translation. Suggestions describes how proposed guidance changes are handled.

To translate content from an application, make your first API request. The recipe takes a small payload through setup, translation, and inspection of a saved target file. If your source already lives in files, the CLI quickstart follows that path through retrieval instead.