Usage of RAG/KB to train not only search

Hi, I’m trying to train a custom GPT (can be Assistant API) in order to understand how to get insights from past surveys (large surveys). The idea is upload to its KB 1 zip file with multiple csv files insides that belong to 1 survey and named is Survey 1 Example.zip, then upload another text file names Survey 1 Insights Example 1.txt and so on for a fuew surveys. Now, when I upload a new zip fiie with a new survey and want my GPT to get insights, how can I expect the GPT to use those example to be trained on how to get insights from a survey and not used like a place to search for text with cosine similarity in a pdf?
Thanks in advance. Guillermo.