Translation

DXPR AI Provider implements the AI module's translate_text operation with a content-aware prompt engine. When you set DXPR as your translation provider, every translation request uses dynamic prompts that adapt to the content being translated instead of a static template.

Setup

Translation requires three modules working together:

1. AI Translate (from the AI module package): provides the translation UI and field-level translation workflow
2. Content Translation (Drupal core): enables per-language content versions
3. DXPR AI Provider: handles the actual translation calls through the Kavya platform

Step 1: Enable AI Translate

drush en ai_translate

This also enables Drupal's Language and Content Translation modules if they are not already active.

Step 2: Add languages

Navigate to /admin/config/regional/language/add and add the languages your site needs. Drupal ships with support for over 100 languages.

Step 3: Set DXPR as the translation provider

Navigate to /admin/config/ai/settings and set DXPR AI Provider as the default provider for the Translate Text operation type. Select the kavya-m1 model (or kavya-m1-eu for EU data residency).

Step 4: Enable translation for content types

Navigate to /admin/config/regional/content-language and enable translation for the content types and fields you want to translate.

Make sure to expand each content type and tick the individual fields (title, body, and any custom fields). Enabling translation at the content type level alone is not enough; unticked fields will be skipped during translation.

Translating content

Once configured, a Translate using [model] link appears for each untranslated language on the content Translations tab (where [model] is the model you selected in step 3):

1. Navigate to a node and open the Translations tab
2. Click Translate using kavya-m1-fast (or whichever model you configured) next to the target language
3. AI Translate sends each field to DXPR AI Provider, which translates it while preserving HTML markup and field structure
4. The translation form opens with all fields pre-filled
5. Review the translation, adjust as needed, and save

Using TMGMT instead

For batch translation workflows, job queues, and translation memory, use TMGMT with the AI Translator plugin:

drush en tmgmt ai_tmgmt

TMGMT provides a translation management dashboard where you can select multiple content items, submit them as a translation job, and review translations before publishing. The AI Translator plugin routes translation requests through the AI module's translate_text operation, so DXPR AI Provider handles the actual translation with the same content-aware prompts.

How the prompts work

DXPR AI Provider builds a different prompt for each translation request based on what the content contains:

• Plain text gets a minimal prompt: translate, maintain tone, adapt cultural references
• HTML content adds rules for preserving markup and translating attributes (alt, title, placeholder, aria-label, and five more)
• RTL target languages (Arabic, Hebrew, Farsi, Urdu, Yiddish, Kurdish, Pashto) add direction-handling instructions
• Prompt injection protection is always included: the model is told to translate only, ignoring any instructions embedded in the source text

The prompt is assembled from these pieces at request time, so plain-text fields never see HTML rules and vice versa. This avoids the "translate everything as if it were HTML" problem of static prompt templates.

Attribute and element coverage

When translating HTML, DXPR AI Provider detects and translates attributes and elements that the default AI Translate fallback skips:

What	DXPR Provider	AI Translate default
Standard attributes	alt, title, placeholder	alt, title, placeholder
Submit/button value	value (on input[submit] and button)	value
Accessibility attributes	aria-label, aria-description, aria-describedby	not translated
UI component attributes	data-tooltip, data-title	not translated
Form elements	label, fieldset, option, optgroup	label, fieldset (title only)
Other elements	abbr, iframe (title)	not translated

Attributes are only added to the prompt when they are actually present in the input, keeping the prompt minimal.

Em-dash handling

The em-dash post-processor resolves its mode based on the target language, not the source. If the target language has a per-language override, that override is used; otherwise the global em-dash mode applies.

This means translations into languages where em-dashes are conventional (such as Russian) can preserve them, while translations into languages where they signal AI authorship (such as English) can replace them.