Our company is planning to use GPT-3.5 Turbo to create our own custom chatbot that is supposed to help the users with their meditation journey. Currently our server is hosted by Microsoft, so we are using their OpenAI Studio to get things going faster. We have a document in Korean with basic information on meditation, like how to start, how often to practice meditation, etc. and we are using hybrid semantic search with text-embedding-ada-002 for vectorizing. When testing out the AI search, the information being retrieved is relevant enough. However the problem occurs when the data is being uploaded to the chat playground and the chat keeps responding with “This question is out of the scope of the retrieved documents. Therefore, I cannot provide an answer based on the documents.” message in English to a prompt written in Korean (same language as the uploaded data). I tried reflecting this in a system message, asking GPT to respond according to its own knowledge to questions for which no information was found in the uploaded document, however that did not work. At certain values of temperature and topP it gives out a relatively good answer, now considering results of the semantic search, but later with the same configurations for both temperature and topP and the exact same prompt, it again responds with “no relevant information was found”. I am not sure what to do at with this issue, so if you could please give any advice I would most certainly appreciate it. In the screenshots provided as you can see same parameters are being used, however one time it responded perfectly while in another case the chatbot apologizes for lack of information.
