Most large language models are trained predominantly on English. Persian — with its right-to-left script, half-spaces, and the gap between formal and colloquial registers — makes up a tiny share of their training data.
You’ve seen the result in practice: grammatically correct answers with a “translated” tone, half-space errors, and weak understanding of local business terminology.
The solution isn’t to discard these models; it’s to adapt them: retrieval over your organization’s own Persian documents (RAG), carefully written Persian instructions, and — for serious cases — fine-tuning on your domain data.
Our experience shows that an assistant connected to internal organizational knowledge can, with today’s models, bring Persian responses to a level where users can barely tell it apart from a human operator.