Adding Amharic and Afaan Oromo
to AI Assistants
Large language models handle Ethiopian languages unevenly. How to evaluate them honestly, handle Ge'ez script and Qubee, and design fallbacks users trust.
Amharic and Afaan Oromo are each spoken by tens of millions of people, yet they are still comparatively low-resource languages for AI. Modern language models can read and write both to a degree, but quality varies widely between models and tasks. Serving these users well takes deliberate engineering, not just a translated system prompt.
Know the scripts
Amharic is written in the Ge'ez (Ethiopic) script, an abugida in which each character represents a consonant-vowel combination. Afaan Oromo is written in Qubee, a Latin-based alphabet in which doubled vowels and consonants change meaning. Both affect how text is tokenised, searched and validated.
- Normalise input: several Ethiopic characters share a sound (for example ሀ, ሐ and ኀ, or ሰ and ሠ) and people type them interchangeably, so search should treat them as equivalent.
- Expect mixed-language messages. People switch between Amharic, Afaan Oromo and English in one conversation, and sometimes write Amharic in Latin letters.
- Check that fonts, input fields and PDF exports render Ethiopic characters correctly on the devices your users actually have.
Evaluate before you promise
Do not assume a model that is strong in English is strong in Amharic. Build a small evaluation set from real, anonymised questions in each language, have native speakers grade the answers, and rerun it whenever you change the model or the prompt.
What to measure per language
| Check | Why it matters | Who judges |
|---|---|---|
| Factual accuracy | Fluent but wrong answers erode trust fastest | Domain expert |
| Fluency and register | Literal or archaic phrasing feels foreign | Native speaker |
| Language consistency | Replies should stay in the user's language | Automated check and reviewer |
| Safety behaviour | Moderation can be weaker outside English | Reviewer |
Ground answers in your own content
Retrieval-augmented generation (RAG) helps a great deal: index your help articles, policies and product data, retrieve the relevant passages and ask the model to answer from them. Keep source documents in the languages you serve where possible, because cross-language retrieval is less reliable, and test your embedding model on Amharic and Afaan Oromo queries specifically.
Design graceful fallbacks
- Let users choose their language explicitly instead of relying only on automatic detection.
- Hand over to a human agent, with the conversation history, when the assistant is unsure.
- For high-stakes answers about money, health or legal matters, show the source passage next to the reply.
Voice is a separate project
Many users prefer voice notes to typing Ge'ez script on a phone keyboard. Speech recognition and text-to-speech quality for Ethiopian languages differs from text quality, so evaluate it separately with recordings that reflect real accents and background noise.
What teams ask before launching an assistant in Ethiopian languages.
Can we fine-tune a model on our own Amharic data?
Yes, but start with prompting and retrieval. Fine-tuning helps with tone and format once you have enough clean, consented examples, and it needs its own evaluation.
Is translating through English a good shortcut?
It can work as a fallback, but it adds latency and loses names, idioms and nuance. Measure it against direct generation on your evaluation set before relying on it.
Where should conversations be stored?
Treat them as personal data: decide where they are stored and for how long, and follow Ethiopia's Personal Data Protection Proclamation and any sector rules that apply to you.
Want to talk to an Arki engineer?
Before any commercial conversation, we'll put you on a 30-min call with a senior engineer in your stack. No salespeople. Just a technical sanity check.
Tell us what you're working on.
We'll respond within one business day with the right next step.