Multilingual content retrieval and answer generation

This is for:

Developer System Administrator

A Coveo Machine Learning (Coveo ML) model that uses a text encoder to process language can be configured to support multiple languages. Currently, this applies to the following Coveo Machine Learning (Coveo ML) models:

By default, these models use an English-only text encoder, and therefore support processing and retrieving only English content. However, Coveo offers beta support for all of the other index-supported languages.

Important

Coveo offers beta support for languages other than English. This means that languages other than English are not fully tested or supported by Coveo and performance may vary.

This article describes how to configure these models to process and retrieve content, and generate content in the case of RGA, in any of the index-supported languages.

Multilingual configuration overview

If you have content in any of the index-supported languages, multilingual configuration allows the Passage Retrieval (CPR), Relevance Generative Answering (RGA), Semantic Encoder (SE), and Smart Snippet models to create the embeddings that the model then uses to retrieve content based on semantic similarity.

For the RGA model, the configuration also includes setting the language for the generated answers. Typically, you’d set it to match the language of the content that the RGA model retrieves, but you can set it to a different language using the generative LLM’s translation capabilities.

Specifically, multilingual configuration allows each of the models to do the following in the chosen language:

  • CPR: Retrieve the most relevant passages.

  • RGA: Generate answers.

  • SE: Retrieve the most relevant content for CPR and RGA during first-stage content retrieval.

    Note

    When using CPR or RGA, you must also configure the SE model to support multilingual content.

  • Smart Snippets: Retrieve the most relevant item and display a snippet of that item in the search results.

To configure multilingual support

Configure multilingual content retrieval

This configuration applies to CPR, RGA, SE, and Smart Snippets, and consists of setting the model to use a multilingual text encoder, and specifying the dataset and dataset languages.

CPR, RGA, SE, and Smart Snippets use a pre-trained text encoder to create the embeddings that the model then uses to retrieve content based on semantic similarity. By default, these models use an English-only text encoder. Configuring a model to use a multilingual text encoder allows the model to effectively create the embeddings for content in different languages, and then retrieve content in the desired language based on those embeddings.

Important

Associate your model with a search interface and query pipeline that are configured to handle multiple languages. If the associated query pipeline doesn’t filter content based on the desired language, the model’s multilingual encoder retrieves content from all available items in its training set based on semantic similarity, regardless of language.

To configure multilingual content retrieval

  1. If you haven’t yet created the model, create the model using the Administration Console.

  2. Access the Machine Learning section of the Swagger UI that’s associated with your Coveo organization region (US | CA | EU | AU).

  3. Click Authorize and authenticate using your Coveo organization account credentials.

  4. In the Swagger UI, expand the Machine Learning Models section.

  5. Use the List all models GET request to access your model’s JSON code:

    Use Swagger UI to get a list of machine learning models | Coveo

    1. Enter the organizationId of your Coveo organization.

    2. Filter the list by entering the engine Id of the model type you want to configure. Use chunksretrieval for CPR, genqa for RGA, embeddings for SE, and mlquestionanswering for Smart Snippets.

    3. Click Execute. The JSON code for all the models of the chosen type in your organization appears in the Response body.

    4. In the Response body, copy the JSON code of the model that you want to configure. You’ll need it in the next step.

  6. Use the Update the configuration of a model PUT request to modify the model’s JSON code:

    Use Swagger UI to edit a machine learning model | Coveo

    1. Paste the copied JSON code in the Request body.

    2. Enter the organizationId of your Coveo organization.

    3. Enter the modelId of the model you want to modify.

      Tip

      If you don’t know the model ID, it appears in the model’s JSON code in the id field.

    4. Modify the JSON code to set the model to use a multilingual text encoder, and specify the dataset and dataset languages:

      1. In the extraConfig element of the JSON, add the encoderConfig parameter to set the multilingual text encoder to MULTILINGUAL_E5_SMALL.

        {
...
"extraConfig": {
    "encoderConfig": {
    "name": "MULTILINGUAL_E5_SMALL"
    },
    }
...
}

      2. In the indexExport element of the JSON, use the query parameter to specify the sources that contain the datasets that you want the model to use, as well as the content languages.

        Tip

        You can set a model to use dataset content in different languages, and then configure the query pipeline to filter the content the model retrieves based on the required language.
        Example

        You want your RGA model to use the datasets from two sources: mysource1 and mysource2. The two sources contain content in English, French, and German, and the languages are specified using the language field in the source items.

        The query parameter would be:

        {
...
"extraConfig": {
    "indexExport": {
      "query": "(@source==(\"mysource1\",\"mysource2\") @language==(English,French,German))"
    },
}
...
}

        If you remove German from the list of languages, the RGA model will use only the English and French content from those sources. The German content will not be embedded, retrieved, or used for answer generation.

      3. In the indexExport element, delete the sources parameter from the JSON, if it exists. The sources are now specified in the query parameter, so the sources parameter is no longer needed.

  7. In the Swagger UI, click Execute to apply the changes to the model. You can review the model information in the Administration Console to see if your changes were applied successfully.

    Use Swagger UI to push edits to a machine learning model | Coveo

Set the language for generated answers

This configuration applies only for RGA. If you’ve configured your RGA model to retrieve content in multiple languages, you can apply additional configuration to set the language of the generated answers. Typically, you’d set it to match the language of the content that the RGA model retrieves, but you can set it to a different language. If set to a different language than the content, the generative LLM translates the generated answer.

To determine the language of the generated answer, the RGA model relies on a locale setting that specifies the language and region. For example, en-US is the locale for English in the United States, and fr-FR is the locale for French in France.

Depending on how you choose to configure your RGA implementation, the locale that’s used for the generated answer can come from the search interface, the query pipeline, or the RGA model configuration itself. Once a locale is established, the RGA model includes that locale in the prompt that’s sent to the LLM for answer generation. The generative LLM then generates the answer in that language.

No matter how you configure your RGA implementation, RGA ultimately relies on two parameters to establish the locale: targetLocaleFromRequest and targetLocale. These parameters are typically set in the RGA model configuration, but they can also be set in the RGA model association in the query pipeline. How and where you set these parameters is based on your chosen strategy for setting the locale.

To set the language used by RGA for generated answers

  1. Configure multilingual content retrieval for the RGA model.

    Note

    Because the SE model is used by RGA to retrieve relevant content, the SE model must also be configured to retrieve multilingual content.

  2. Determine your locale strategy to know where and how to set the locale that’s used for generated answers. This is based on your implementation needs and RGA’s locale workflow, and may require you to do one or both of the following:

  3. Set the RGA locale parameters (targetLocaleFromRequest and targetLocale) that RGA will use to determine the language for generated answers.

Determine your locale strategy

Before we get into specifics on how to configure the locale for your RGA implementation, let’s first look at where the locale can be set in the RGA workflow. For maximum flexibility, the locale that’s used for generated answers can come from the search interface, the query pipeline that’s used by the RGA model, or from the RGA model configuration itself.

The following diagram illustrates the path a locale setting takes in an RGA workflow, beginning with the recommended locale setting in the search interface. This is not to say that the locale setting must absolutely be set in the search interface. It can be set at any of the indicated points along the path, with each subsequent step having the potential to override the locale set in a previous step. You can use this diagram to help you determine the optimal implementation method for your specific needs.

Important

Associate your model with a search interface and query pipeline that are configured to handle multiple languages. If the associated query pipeline doesn’t filter content based on the desired language, the model’s multilingual encoder retrieves content from all available items in its training set based on semantic similarity, regardless of language.
RGA answer generation locale flow | Coveo

1

(Optional) If the locale is set in your RGA-enabled search interface, the locale value is sent as part of the query request to the query pipeline that’s used by the RGA model.

Note

Coveo recommends that you use the search interface to set the locale to use for generated answers. It’s the most reliable way to use your website’s existing business logic, and provide a consistent customer experience without introducing unnecessary complexity in your RGA implementation.

2

(Optional) You can create a query parameter rule in the query pipeline to override the locale value from the search interface, or to set the locale value. If you don’t set a query parameter rule for the locale, the locale that’s received from the search interface is kept.

3

(Required) The targetLocaleFromRequest and targetLocale parameters are ultimately what determine the locale the RGA model uses for the generated answer. Based on the parameter settings, the RGA model uses either the locale it receives from the search interface or query parameter rule, or the locale that’s set in targetLocale. For increased flexibility, the parameters can be set either in the RGA model association in the query pipeline, in the RGA model configuration itself, or in both. If set in both, the parameter settings in the RGA model association take precedence. For details, see Set the RGA locale parameters.

Note

Unlike 1 and 2 that are optional, you must set the targetLocaleFromRequest and targetLocale parameters. Otherwise, the RGA model will always default to English for generated answers, no matter what locale is set in a previous step.
Tip

See example configurations of how to configure your locale settings based on different scenarios.

Set the locale in the search interface

While setting the locale in your RGA-enabled search interface is optional, it’s the recommended way to set the locale for generated answers.

Note

Even if you set the locale in the search interface, you must still set the RGA locale parameters.

Enterprises typically build their websites and search interfaces to serve specific business logic and audiences. It’s for this reason that Coveo recommends that you set the locale in the search interface to take advantage of the same business logic when it comes to the language of generated answers. Allowing the search interface to dictate the language for generated answers is the most reliable way to provide a consistent customer experience without introducing unnecessary complexity in your RGA implementation.

Tip

Even if you configure a locale for all your search interfaces, you can still override the locale later in the RGA workflow to accommodate for specific use cases and flows. In other words, you could use the search interface locale for the majority of use cases, but override it when needed.

Set the locale for your RGA-enabled search interface based on the library or framework that you used to build your search interface:

Important

The locale used for RGA must be a valid tag that includes a language code, or language and region codes, according to the IETF BCP 47 standard. For example, use fr-FR for French in France or de for international German. See Supported locale tags for details.

Set the locale using a query parameter rule

There may be instances where you want to use a query parameter rule in the query pipeline to set or override the locale value that the RGA model receives. This is especially useful if you’re using the same query pipeline for multiple search interfaces, or if you want to set the locale based on a specific condition.

Note

Even if you set the locale using a query parameter rule, you must still set the RGA locale parameters.

To create a query parameter rule for the locale

Add a query parameter rule using the Locale parameter in the query pipeline that’s used by your RGA-enabled search interface.

Important

The locale used for RGA must be a valid tag that includes a language code, or language and region codes, according to the IETF BCP 47 standard. For example, use fr-FR for French in France or de for international German. See Supported locale tags for details.

Set the RGA locale parameters

The RGA model references the targetLocaleFromRequest and targetLocale parameter settings to determine what locale to use for the generated answer.

Important

These parameters are required to generate answers in a language other than English. If the parameters are not set, the generated answers will always default to English.

  • targetLocaleFromRequest: Sets whether to ignore or use the locale that the RGA model receives with the query request. It can be set to one of the following values:

    • ignore: Instructs the RGA model to ignore any locale that it receives, and instead use the locale that’s specified in the targetLocale parameter.

    • fallback_if_invalid: Instructs the RGA model to use the locale that it receives. In the case of an invalid locale, the RGA model will use the locale that’s specified in the targetLocale parameter.

  • targetLocale: Sets the locale to use if targetLocaleFromRequest is set to ignore, or if it’s set to fallback_if_invalid and the locale is invalid.

    Important

    The locale used for RGA must be a valid tag that includes a language code, or language and region codes, according to the IETF BCP 47 standard. For example, use fr-FR for French in France or de for international German. See Supported locale tags for details.
Tip

See example configurations of how to configure your locale settings based on different scenarios.

To set the RGA locale parameters

You can set the targetLocaleFromRequest and targetLocale parameters in the RGA model association in the query pipeline, or in the RGA model configuration. For more information, see Model association versus model configuration.

Set the RGA locale parameters in the model configuration

Set the RGA locale parameters in the model configuration

  1. Access the Machine Learning section of the Swagger UI that’s associated with your Coveo organization region (US | CA | EU | AU).

  2. Click Authorize and authenticate using your Coveo organization account credentials.

  3. In the Swagger UI, expand the Machine Learning Models section.

  4. Use the List all models GET request to access your model’s JSON code:

    1. Enter the organizationId of your Coveo organization.

    2. Filter the list to show only RGA models by entering genqa for the engine Id.

    3. Click Execute. The JSON code for all your RGA models appears in the Response body.

    4. In the Response body, copy the JSON code of the model that you want to configure. You’ll need it in the next step.

  5. Use the Update the configuration of a model PUT request to modify the model’s JSON code:

    1. Paste the copied JSON code in the Request body.

    2. Enter the organizationId of your Coveo organization.

    3. Enter the modelId of the model you want to modify.

      Tip

      If you don’t know the model ID, it appears in the model’s JSON code in the id field.

    4. In the extraConfig element of the JSON, add the targetLocaleFromRequest and targetLocale locale parameters with your chosen settings.

      Important

      The locale used for RGA must be a valid tag that includes a language code, or language and region codes, according to the IETF BCP 47 standard. For example, use fr-FR for French in France or de for international German. See Supported locale tags for details.
      Example

      You want your RGA model to use the locale that it receives from the search interface, and use English as the fallback language if the locale is invalid.

      The targetLocaleFromRequest and targetLocale parameter settings would be:

      {
...
"extraConfig": {
    "encoderConfig": {
    "name": "MULTILINGUAL_E5_SMALL"
    },
    "indexExport": {
      "query": "(@source==(\"mysource1\",\"mysource2\") @language==(EN,FR,DE))"
    },
    "targetLocaleFromRequest": "fallback_if_invalid",
    "targetLocale": "en-US"
},
...
}

  6. In the Swagger UI, click Execute to apply the changes to the model. You can review the model information in the Administration Console to see if your changes were applied successfully.

Set the RGA locale parameters in the model association in the query pipeline

Set the RGA locale parameters in the model association

  1. On the Query Pipelines (platform-ca | platform-eu | platform-au) page, click the query pipeline that’s used by your RGA-enabled search interface, and then click Edit components in the Action bar.

  2. Select the Machine Learning tab.

  3. Double-click the RGA model association that you want to edit.

  4. If the Edit a Model Association subpage opens in JSON view, proceed to the next step. Otherwise, in the upper-right corner, click dots, click Switch to JSON view.

  5. In the customQueryParameters element of the JSON, add the targetLocaleFromRequest and targetLocale locale parameters with your chosen settings.

    Important

    The locale used for RGA must be a valid tag that includes a language code, or language and region codes, according to the IETF BCP 47 standard. For example, use fr-FR for French in France or de for international German. See Supported locale tags for details.
    Example

    You want your RGA model to ignore the locale that it receives, and use French for generated answers.

    The targetLocaleFromRequest and targetLocale parameter settings would be:

    {
...
"customQueryParameter": {
    ...
    "targetLocaleFromRequest": "ignore",
    "targetLocale": "fr-FR"
},
...
}

Model association versus model configuration

You can set the targetLocaleFromRequest and targetLocale parameters in the RGA model association in the query pipeline, or in the RGA model configuration itself.

Note

Depending on your implementation needs, you can also set the parameters in both the RGA model association and RGA model configuration. If the parameters are set in both, the settings in the model association take precedence and the ones in the model configuration are ignored.

Depending on your implementation needs, you can set the parameters as follows:

  • Both parameters set in the model association only.

  • Both parameters set in the RGA model configuration only.

  • One parameter setting in the model association and another in the model configuration. This is an advanced configuration, but it’s useful in situations where you use the same RGA model across multiple query pipelines, or if you use the same RGA model more than once in the same query pipeline. In this case, a parameter setting in the model association takes precedence over the setting in the model configuration. If a parameter is set in the model association, the corresponding setting in the model configuration is ignored.

    Example

    You want to use the same RGA model in three different query pipelines. One query pipeline is used for English content, and the other two are used for French and German content respectively. In each case, you want the RGA model to ignore the locale it receives, and use a different locale for each query pipeline.

    You can set the parameters as follows:

    • In the RGA model configuration, set the targetLocaleFromRequest parameter to ignore.

    • In the RGA model association in the query pipeline, set the targetLocale parameter to en-US, fr-FR, or de-DE depending on the query pipeline.

    Given this configuration, since the targetLocaleFromRequest parameter is not set in the model association, it uses the setting in the RGA model configuration, and the targetLocale setting in each of the model associations is used. If the model configuration had a targetLocale setting, it would be ignored in favor of the setting in the model association since that setting takes precedence.

Leverage the LLM translation capabilities

It’s possible to generate answers in a language that’s different from the language of the content that the RGA model retrieves.

RGA leverages GPT’s linguistic capabilities to generate the answer. This includes the LLM’s ability to translate text.

During answer generation, the RGA model sends a prompt that includes the retrieved chunks and the desired locale to the GPT LLM that’s used to generate the answer. If the language of the chunks is different from the requested locale language, the GPT LLM generates the answer using the original content and then translates the answer before streaming it back to the search interface.

When the generated answer is in a different language than the content that was used to generate it, the citations for the generated answer appear in the original language.

Example

If the RGA model retrieves content in English and the locale for answer generation is set to Japanese (ja-JP), the answer appears in Japanese, but the citations are for the English content.

RGA answer translation example | Coveo
Important

RGA doesn’t control the quality of the translation. The GPT LLM translates the answer on-the-fly, and the translated chunks are not stored in the Coveo index.

While the GPT LLM’s translation capabilities may be useful to bridge content gaps in certain languages, you should use caution for content that requires high-quality translations, such as legal or technical documents.

Example configurations

The following examples show how to configure the locale for generated answers in different scenarios.

Generate answers in a single language that is not English

Unilingual workflow - Locale set in the RGA model configuration

Scenario:

  • Your enterprise has a French website that uses a search interface with RGA.

  • Your dataset includes only French content.

  • You have one RGA model that’s used for your French dataset.

Configuration:

  • In your RGA model configuration, set targetLocaleFromRequest to ignore, and targetLocale to fr-FR.

    {
...
"extraConfig": {
    ...
    "targetLocaleFromRequest": "ignore",
    "targetLocale": "fr-FR"
},
...
}

  • Assign the RGA model to the query pipeline that’s used for the search interface.

Result:

This sets the RGA model to ignore any locale it receives and always use fr-FR as the language for generated answers.

Generate answers using the language of my search interface

Multilingual workflow with separate query pipelines - Locale set in the RGA model configuration

Scenario:

  • Your enterprise has English, French, and Spanish versions of its website, and each website includes a search interface that uses RGA.

  • Your dataset includes content in English, French, and Spanish.

  • You have one RGA model that’s used for all of your dataset content (English, French, and Spanish).

  • You use a separate query pipeline for each of your search interfaces. Each query pipeline filters the content to use based on the corresponding search interface. For example, for the French search interface, the associated query pipeline filters the content to use only French content.

Configuration

  • Set the locale in the search interface based on the language of the website: en-EN for English, fr-FR for French, and es-ES for Spanish.

  • In your RGA model configuration, set targetLocaleFromRequest to fallback_if_invalid, and targetLocale to en-US.

    {
...
"extraConfig": {
    ...
    "targetLocaleFromRequest": "fallback_if_invalid",
    "targetLocale": "en-US"
},
...
}

  • Your RGA model is associated with each of your query pipelines. In the query pipeline that’s used for the English search interface, filter the content so that only English content is used. Do the same for the query pipelines for French and Spanish, but filter the content to French and Spanish respectively.

Result:

  • Each search interface sends its locale with the query request to the corresponding query pipeline.

  • For the English search interface, the RGA model receives en-EN as the locale and generates the answer in English using the English content only.

  • For the French search interface, the RGA model receives fr-FR as the locale and generates the answer in French using the French content only.

  • For the Spanish search interface, the RGA model receives es-ES as the locale and generates the answer in Spanish using the Spanish content only.

  • If the locale the RGA model receives is invalid, the fallback language is en-US.
Multilingual workflow with one query pipeline - Locale set in the RGA model associations

Scenario:

  • Your enterprise has English, French, and Spanish versions of its website, and each website includes a search interface that uses RGA.

  • Your dataset includes content in English, French, and Spanish.

  • You have one RGA model that’s used for all of your dataset content (English, French, and Spanish).

  • You use one query pipeline for all of your search interfaces. The same RGA model is associated with the query pipeline three times, with each association having a condition that links it to a different search interface.

Configuration:

  • Set the locale in the search interface based on the language of the website: en-EN for English, fr-FR for French, and es-ES for Spanish.

  • In the query pipeline, add a filter rule to filter the content that’s used based on a language condition.

  • For each of your RGA model associations in the query pipeline, set targetLocaleFromRequest to ignore and targetLocale to the locale of the search interface. For example, in the RGA model association that’s used for your Spanish search interface, set targetLocale to es-ES.

    {
...
"customQueryParameter": {
    ...
    "targetLocaleFromRequest": "ignore",
    "targetLocale": "es-ES"
},
...
}

Result:

  • Each search interface sends its locale with the query request to the query pipeline.

  • The targetLocaleFromRequest and targetLocale parameter settings in the respective RGA model associations take effect.

    Note

    Even if the RGA model configuration had targetLocaleFromRequest and targetLocale parameters set, the settings in the model association take precedence and the settings in the model configuration are ignored.

  • For each of the search interfaces, the language used for the generated answers is based on the targetLocale setting in the corresponding model association. The filter added to the query pipeline ensures that only the content in the corresponding language is used.

    Tip

    Since the locale is set in the search interface, and the locale is sent with the query request, you can also set targetLocaleFromRequest to fallback_if_invalid and targetLocale to en-US in each of the model associations. This results in the same behavior except that English is used as the fallback language if the locale received is invalid.

Generate answers in the language of my search interface with content that might be in a different language

Multilingual workflow using generative LLM translation capabilities

Scenario:

  • Your enterprise has English and French versions of its website, and each website includes a search interface that uses RGA.

  • Your dataset includes content in English only.

  • You have one RGA model that’s used for all of your English dataset content.

  • You use a separate query pipeline for each of your search interfaces. The RGA model is associated with each query pipeline.

Configuration:

  • Set the locale in the search interface based on the language of the website: en-EN for English and fr-FR for French.

  • In the RGA model configuration, set targetLocaleFromRequest to fallback_if_invalid and targetLocale to the en-US of the search interface.

    {
...
"extraConfig": {
    ...
    "targetLocaleFromRequest": "fallback_if_invalid",
    "targetLocale": "en-US"
},
...
}

Result:

  • Each search interface sends its locale with the query request to the query pipeline.

  • For the English search interface, the RGA model receives en-EN as the locale and generates the answer in English using the English content.

  • For the French search interface, the RGA model receives fr-FR as the locale. The answer is generated using the English content, but the answer is translated to French by the generative LLM. Though the generated answer appears in French, the citations appear in English.

    Tip

    Since the locale is set in the search interface, and the locale is sent with the query request, you could have set the locale parameters in the query pipeline RGA model association configuration instead. Setting targetLocaleFromRequest to ignore and targetLocale to either en-US or fr-FR depending on the query pipeline results in the same outcome.

Supported locale tags

The locale that’s used by RGA for answer generation must be a valid tag that includes the language, or language and region, as per the IETF BCP 47 standard (language code or language code-region code).

The language code must follow the ISO 639-1 standard, which provides two-letter codes for languages, such as en for English and fr for French.

The region code must follow the ISO 3166-1 standard, such as FR for France and US for United States.

The following table lists the most common locale tags that you can use for multilingual content retrieval and answer generation:

Language Region Locale

Arabic

International

ar

Algeria

ar-DZ

Egypt

ar-EG

Morocco

ar-MA

Saudi Arabia

ar-SA

Chinese

International

zh

Mainland China, simplified characters

zh-CN

Hong Kong, traditional characters

zh-HK

Singapore

zh-SG

Taiwan, traditional characters

zh-TW

Croatian

International

hr

Croatia

hr-HR

Czech

International

cs

Czechia

cs-CZ

Danish

International

da

Denmark

da-DK

Dutch

International

nl

Belgium

nl-BE

Netherlands

nl-NL

English

International

en

Australia

en-AU

Canada

en-CA

United Kingdom

en-GB

New Zealand

en-NZ

United States

en-US

Estonian

International

et

Estonia

et-EE

Finnish

International

fi

Finland

fi-FI

French

International

fr

Belgium

fr-BE

Canada

fr-CA

Switzerland

fr-CH

Algeria

fr-DZ

France

fr-FR

German

International

de

Austria

de-AT

Switzerland

de-CH

Germany

de-DE

Luxembourg

de-LU

Greek

International

el

Cyprus

el-CY

Greece

el-GR

Hebrew

International

he

Israel

he-IL

Hungarian

International

hu

Hungary

hu-HU

Italian

International

it

Switzerland

it-CH

Italy

it-IT

Japanese

International

ja

Japan

ja-JP

Korean

International

ko

South Korea

ko-KR

Norwegian

International

no

Norway

no-NO

Polish

International

pl

Poland

pl-PL

Portuguese

International

pt

Angola

pt-AO

Brazil

pt-BR

Portugal

pt-PT

Romanian

International

ro

Moldova

ro-MD

Romania

ro-RO

Russian

International

ru

Belarus

ru-BY

Kazakhstan

ru-KZ

Russia

ru-RU

Serbian

International

sr

Cyrillic, Serbia

sr-Cyrl-RS

Slovak

International

sk

Slovakia

sk-SK

Slovenian

International

sl

Slovenia

sl-SI

Spanish

International

es

Argentina

es-AR

Colombia

es-CO

Spain

es-ES

Mexico

es-MX

Peru

es-PE

Swedish

International

sv

Sweden

sv-SE

Turkish

International

tr

Cyprus

tr-CY

Turkey

tr-TR

Ukrainian

International

uk

Ukraine

uk-UA