Arabic Support

Overview#

Anar is built Arabic-first. Every product handles Arabic text correctly -- from right-to-left rendering in dashboards to dialect-aware speech transcription in Voice. The platform supports Modern Standard Arabic (MSA) and five regional dialect families across the GCC.

Arabic-Capable Models#

The platform leverages models with strong Arabic support:

Choosing a Model for Arabic Tasks#

• General Arabic NLP (summarization, classification, Q&A): allam-2-7b or llama-3.3-70b-versatile
• Sovereign Arabic workloads: gpt-4o-uaenorth (data stays in UAE)
• Arabic speech-to-text: whisper-large-v3
• Arabic embeddings/RAG: text-embedding-3-large-uaenorth

Dialect Detection#

Anar Voice identifies five Arabic dialect families:

Voice products detect the dialect during transcription and include it in the response metadata. This enables downstream products to adjust their processing -- for example, translation can account for dialect-specific vocabulary.

RTL Support in Dashboards#

All dashboard frontends support bidirectional text rendering:

CSS Direction#

Apply dir="rtl" on Arabic content containers:

<div dir={language === "ar" ? "rtl" : "ltr"} className="font-arabic">
  <p>{arabicContent}</p>
</div>

Font Stack#

The design system defines a dedicated Arabic font family:

--font-arabic: 'IBM Plex Sans Arabic', 'Noto Sans Arabic', sans-serif;

Use font-arabic in Tailwind classes for Arabic text:

<h2 className="font-arabic text-2xl">
  {/* Arabic heading */}
</h2>

Layout Considerations#

• Flexbox: Use flex-row-reverse or CSS logical properties (margin-inline-start instead of margin-left)
• Icons: Directional icons (arrows, chevrons) should flip in RTL mode
• Tables: Align numeric columns to the left even in RTL layouts
• Forms: Input fields should use text-align: right for Arabic input

Arabic OCR (Docs)#

The Docs product uses a three-tier OCR fallback chain optimized for Arabic:

1. Azure Computer Vision -- best accuracy for printed Arabic
2. EasyOCR -- strong character-level Arabic recognition
3. Tesseract -- fallback with Arabic language pack

Configure OCR languages:

OCR_LANGUAGES=ar,en

Results include confidence scores and detected language per page, so mixed Arabic/English documents are handled correctly.

Arabic Translation (Translate)#

The Translate product specializes in Arabic-English document translation with domain awareness:

Translation Memory#

The translation memory (TM) stores previously translated segments and uses fuzzy matching to suggest translations for similar text. TMX import/export is supported for interoperability with CAT tools:

# Import a TMX file
curl -X POST http://localhost:8012/api/v1/tm/import \
  -F "file=@translations.tmx"

# Export translation memory
curl http://localhost:8012/api/v1/tm/export?format=tmx > memory.tmx

Glossary Management#

Define domain-specific glossary entries to ensure consistent terminology:

curl -X POST http://localhost:8012/api/v1/glossaries \
  -H "Content-Type: application/json" \
  -d '{
    "source_term": "artificial intelligence",
    "target_term": "الذكاء الاصطناعي",
    "domain": "technical"
  }'

Arabic Presentations (Present)#

Present generates slides with full Arabic RTL support:

• Slide text renders right-to-left
• The Midnight Gold theme applies consistently to Arabic layouts
• PPTX and PDF exports preserve RTL formatting
• Mixed Arabic/English slides handle bidirectional text correctly

Arabic Meeting Intelligence (Minutes)#

Minutes provides six meeting template categories with Arabic support:

• Templates include Arabic-language fields for agenda, action items, and decisions
• Audio transcription uses Whisper with Arabic dialect detection
• Summaries are generated in the meeting's primary language
• Action items are extracted from both Arabic and English discussions

Bilingual Content Handling#

Many GCC government documents contain mixed Arabic and English text. The platform handles this at multiple levels:

Document Processing#

Docs detects the primary language per page and applies the appropriate OCR model. Mixed-language pages use combined processing with language-tagged output.

RAG and Search#

Chat's RAG pipeline stores language metadata per chunk. Queries in Arabic retrieve Arabic chunks, and cross-language retrieval is supported through multilingual embeddings.

Guard Content Scanning#

Guard's content filtering includes bilingual toxic keyword detection and restricted topic lists in both Arabic and English. PII detection patterns cover GCC-specific formats (UAE national IDs, Saudi Iqama numbers, GCC phone formats).