When Building An LLM Application That Responds In Korean, The Korean Prompts Do Not Perform Well At The GPT 3.5 Level. It Is Difficult To Control The Output Of LLM With Simple Korean Prompts, Especially When You Need To Respond With Honorifics(존댓말, 높임말) In A RAG Implementation. In This Case, The Following Prompts, Which Use A Combination Of English And Korean, Work Well.

When building an LLM application that responds in Korean, the Korean prompts do not perform well at the GPT-3.5 level. It is difficult to control the output of LLM with simple Korean prompts, especially when you need to respond with honorifics(존댓말, 높임말) in a RAG implementation. In this case, the following prompts, which use a combination of English and Korean, work well.