add follow up question recommendations

[tatakof]

use 'gpt-3.5-turbo-0125', ask it to generate 3 questions based on the initial question and the response (and the top_k?)

[tatakof]

follow up questions should be created with only 1 api call and should be extracted either with a fuzzy search (because the structure might not be consistent) or with 'gpt-3.5-turbo-instruct'

[tatakof]

evaluate if the question generator should use some form of memory (compressed llm-output) instead of zero-shoting everything