Smartling Language Adaptation is a fully managed, LLM-powered adaptation Profile that updates regional terminology, spelling, and conventions for locale pairs that share the same language. It replaces the legacy Prompsit adaptation engine.
Smartling Language Adaptation is Smartling's fully managed, LLM-powered adaptation Profile designed for teams serving multiple regional markets that share the same language but use different locales, such as English (United States) and English (United Kingdom). Instead of requiring a full retranslation, Language Adaptation uses an LLM with prompts specifically optimized for regional adaptation. Unlike standard translation prompts, which may unnecessarily rephrase or restructure content, adaptation preserves existing content wherever possible and changes only what is needed for the target locale.
Language Adaptation updates regional terminology, spelling, and conventions while keeping the rest of the content intact. Smartling manages the prompts, model selection, and configuration settings, so no setup or LLM expertise is required.
Smartling Auto Select LLM also handles language adaptation automatically. When source and target locales share the same language, Auto Select LLM detects this and applies an adaptation-specific prompt. You only need a separate Language Adaptation Profile if you want to configure different RAG settings or linguistic packages for adaptation work with non-project-based integrations.
Language Adaptation replaces the legacy Prompsit adaptation engine with broader language support and higher quality across all tested locale pairs.
How it works
Language Adaptation uses leading models from Vertex AI (Gemini), OpenAI, and Amazon Bedrock. Smartling continuously benchmarks against new models as they are released and updates the model as higher-quality options become available, so you are never locked into a single model or provider.
The Profile also includes fallback models. If the primary model is unavailable, Smartling automatically routes the request to a backup model to ensure uninterrupted adaptation.
Retrieval-Augmented Generation (RAG)
Language Adaptation supports RAG by default. RAG is automatically enabled and enhances adaptation quality by referencing the following linguistic assets from your project:
- Translation Memory Examples
- Glossary Terms
- Style Rules for AI
You can disable any of these linguistic assets using the checkboxes in the Profile settings under Assets References.
Set up Language Adaptation
A Language Adaptation Profile is pre-provisioned in your account. You can find it by going to AI Hub > Profiles.
Edit the pre-provisioned Profile
You can edit the existing Language Adaptation Profile to configure RAG settings and linguistic packages for non-project-based integrations (also known as MT integrations) such as the MT API or Smartling Translate:
- Go to AI Hub > Profiles.
- Click the Smartling Language Adaptation Profile name to open its settings.
- Optionally, select Disable hallucination detection to turn off hallucination detection for this Profile. Hallucination detection is enabled by default and is recommended.
- Under Assets References, select which linguistic assets are used for RAG: Translation Memory Examples, Glossary Terms, and Style Rules for AI. All three are enabled by default.
- Select one or more Linguistic Packages to define which linguistic assets (translation memory, glossary, and style guide) are used for RAG with non-project-based integrations. For workflow translations within a project, the project's linguistic package is used by default.
- Click Test Integration to confirm the connection is working. To test the actual adaptation output, you will need to create a test job and submit content for adaptation using the Profile.
- Click Save.
Create an additional Profile
You can create additional Language Adaptation Profiles if needed. For example, you may want separate Profiles to use different linguistic packages for non-project-based integrations.
- Go to AI Hub > Profiles.
- Click Create Profile and select Smartling Language Adaptation.
- Enter a Profile Name.
- Optionally, select Disable hallucination detection to turn off hallucination detection for this Profile. Hallucination detection is enabled by default and is recommended.
- Under Assets References, select which linguistic assets are used for RAG: Translation Memory Examples, Glossary Terms, and Style Rules for AI. All three are enabled by default.
- Select one or more Linguistic Packages to define which linguistic assets are used for RAG with non-project-based integrations such as the MT API or Smartling Translate. For workflow translations within a project, the linguistic package in the project is used by default.
- Click Test Integration to confirm the connection is working. To test the actual adaptation output, create a test job and submit content for adaptation using the Profile.
- Click Save.
Considerations
Cost
For customers with a Smartling AI Subscription, Language Adaptation usage is billed as AI Hub words at the Smartling key rate. If you do not have an AI subscription, usage counts as MT words.
Language support
Language Adaptation supports all Smartling-supported locales, but quality may vary for lower-resource locales. The following locale pairs have been tested by Smartling and deliver better results compared to the legacy Prompsit adaptation engine.
- English (US) [en-US] ↔ English (UK) [en-GB]
- Spanish (Spain) [es-ES] ↔ Spanish (Latin America) [es-LA]
- Spanish (Spain) [es-ES] ↔ Spanish (Mexico) [es-MX]
- Spanish (Spain) [es-ES] ↔ Spanish (US) [es-US]
- French (France) [fr-FR] ↔ French (Canada) [fr-CA]
- Portuguese (Brazil) [pt-BR] ↔ Portuguese (Portugal) [pt-PT]
Migrating from Prompsit
Language Adaptation replaces the legacy Prompsit adaptation engine. Prompsit supported only five locale pairs and used a static, rule-based approach. Language Adaptation supports all locale pairs and has been benchmarked at higher quality across all Prompsit-supported languages. If you currently use Prompsit, update your workflows to use the Language Adaptation Profile. If your organization cannot use LLMs, you will need to use a Copy Source to Target workflow with an additional step for human adaptation as an alternative.
Fine-tuning and custom models
Language Adaptation uses generic models. Fine-tuned or custom models are not supported.
Bring Your Own Key (BYOK)
BYOK is not supported for Language Adaptation. The Profile is provisioned by Smartling and uses a Smartling key.
Hallucination detection and human review
Smartling includes hallucination detection to flag unexpected changes in LLM adaptation output. While this reduces risk, LLM-based adaptation may still occasionally produce unexpected results. A human review step in your workflow is recommended. You can also set up a dynamic workflow using the "Translation matches source" condition to route strings that may need human review to confirm that adaptation was applied correctly and that any glossary terms or translation memory references were adapted as expected.
FAQ
Do I need a separate Language Adaptation Profile if I already use Auto Select LLM?
Not necessarily. Auto Select LLM automatically detects locale pairs that share the same language and applies an adaptation-specific prompt. A separate Language Adaptation Profile is only needed if you want to configure different RAG settings or use different linguistic packages for adaptation work with non-project-based integrations.
What is the difference between Prompsit and Smartling Language Adaptation?
Prompsit is a third-party adaptation engine that uses a static, rule-based system. Language Adaptation is Smartling's own adaptation solution, using LLMs with a prompt specifically tuned for the adaptation use case.
Is Language Adaptation better quality than Prompsit?
Yes. Smartling benchmarked adaptation quality across all Prompsit-supported locale pairs and Language Adaptation outperformed Prompsit in every case. Additionally, Language Adaptation supports all locale pairs, whereas Prompsit supported only five.
What providers and models does Language Adaptation use?
Smartling uses a primary model selected for the best cost-quality ratio, with fallback models if the primary model is unavailable. The model is chosen from leading providers such as Vertex AI (Gemini), OpenAI, and Amazon Bedrock. Smartling benchmarks against new models as they are released and updates the selection when a better option is available. See How it works for more detail.
Is the adaptation prompt visible or customizable?
No. The prompt is not visible and cannot be customized. You cannot add your own prompt messages or modify LLM parameters. If you need direct control over the prompt, create a standard LLM Profile for a specific LLM provider.
What is the difference between Language Adaptation and LLM Translation?
Language Adaptation is designed to make only the minimum changes necessary for the target locale, such as terminology, spelling, and regional conventions, while keeping the source content as intact as possible. LLM Translation, by contrast, produces a full translation and may rephrase or restructure content more broadly. Use Language Adaptation for locale pairs that share a common base language (e.g., en-US to en-GB, or fr-FR to fr-CA).
Can I restrict which LLM providers are used?
Yes. If due to company policy or preference you do not wish the Language Adaptation Profile to use a particular LLM provider, you can restrict the use of an LLM provider within the AI Hub Provider settings.