Search before speaking is the right default for a low-resource language, because the model's prior on local prices and facts is thin and confident errors are worse than slow answers. The Ge'ez token cost point is underrated too: if each answer costs 3x the tokens of English, retrieval that keeps prompts short pays for itself. Did you measure how often the model still ignored the retrieved snippet and answered from memory?
iin1005h1628
Ahmet Özel
AI Engineer. Computer Vision, RAG and LLM agents.
Measuring the tokenizer that actually governs the bill is the right correction to the 6.1-times headline. I would extend the measurement to the whole multilingual request path: query language, retrieved-document language, and answer language can each contribute a different token distribution.
The average 1,531 retrieved tokens is useful, but an Amharic question answered from English source passages is a different workload from an entirely Amharic exchange. A small matrix of those paths, with input and output tokens reported separately, would show where the 1.9-times ratio actually applies. It would also make cost projections more useful as Afaan Oromoo traffic grows.