How we built Bini, an assistant that answers Ethiopians' everyday questions in Amharic, and what we learned about the real cost of the Ge'ez script.
Ask a general-purpose AI assistant how much it costs to renew an Ethiopian passport, or how to recover a lost Fayda number, and you will usually get a confident answer. Often it is out of date, and sometimes it is simply made up. The model learned about Ethiopia from whatever happened to be on the internet when it was trained, and Ethiopian government procedures change faster than those snapshots.
That was the problem we set out to solve at BinaSmart, an Amharic-first app for everyday life in Ethiopia. People come to us to book a ride in Addis Ababa or find a hotel, but they also ask practical questions: what documents a business licence needs, what the new salary tax bands are, which office handles an urgent passport. We wanted an assistant that could answer those questions in Amharic and be right.
Not a chatbot, a librarian
Bini, our assistant, doesn't answer from memory. On every question it searches an indexed library of Ethiopian documents (laws, proclamations, government service pages and official fee tables), reads the passage that matches, and answers from that passage. Every answer has a source you can open and check.
This design choice matters more than the choice of model. When the salary tax bands changed under Proclamation 1395/2025, we didn't need to retrain anything. We added the new proclamation to the library, and Bini started answering from it the same day.
On our public test of 138 real questions about government services, Bini currently answers 96% correctly. The questions it misses are useful too: each one points to a document we haven't indexed yet.
The Amharic token tax, measured
AI is billed in tokens, and Amharic is expensive in tokens. Our own news desk first measured Amharic at 6.1 times the tokens of the same text in English, using OpenAI's o200k_base tokenizer.
BinaSmart runs on Gemini, so we measured again with the tokenizer that actually governs our bill. The result was 1.9 times, not 6.1.
Both numbers are correct, and the gap is the lesson: the penalty belongs to the tokenizer, not to the Ge'ez script. A tokenizer that has learned Amharic splits it into sensible pieces, and one that hasn't breaks it into fragments. On our own text, one token carries 3.55 characters of English, 2.67 of Afaan Oromoo and 1.85 of Amharic.
If you are building for Ethiopian languages, measure with the tokenizer you will actually pay for. Published comparisons may be describing a different model.
Why "just paste everything in" doesn't work
With context windows of a million tokens, it is tempting to skip search entirely and paste the whole library into every prompt.
Our library now measures about 2.47 million tokens, so a one-million-token window holds well under half of it. It has also grown fast, from 3,937 chunks to 8,510 since we first published these figures.
Retrieval solves both problems. Across real questions in Amharic and Afaan Oromoo, Bini sends an average of 1,531 tokens per question, roughly 1,600 times less than the whole library. As the library grows, the cost per answer barely moves.
What we would tell other builders
- Ground answers in documents, not in model memory. For government procedures, fees and laws, being out of date is the most common way to be wrong.
- Show the source. People trust an answer they can check, especially when money or legal status depends on it.
- Measure your language's real cost. Tokenizers differ, and the difference can be more than threefold.
- Build the test set early. Our 138 questions tell us exactly where the gaps are.
Bini now speaks Amharic, English and Afaan Oromoo, and we are adding more Ethiopian languages. The full measurement, including how we count tokens and what we own versus license, is published at bina.et/amharic-ai.
You can try Bini free at bina.et/ai, in the browser or inside Telegram, with no download needed.
Ibrahim Kedir Bedru is the founder of BinaSmart (bina.et), based at JJ Darule Building in Addis Ababa.
Originally published on Medium.
Top comments (0)