Perform topic labeling and key words extraction

[Azzedde]

Because the output of the LLM is not 100% controllable, we can find some entries that are not empty, but have some unwanted behaviour such as this entry:

    {
        "question": "What is the official abbreviation of the project called ',',...?",
        "answer": "I don't really know !"
    }

So we need to implement a topic labeling workflow to see 'semantically' what's happening in our data.
We want to use BERTopic in a first step, then use a very cheap LLM that also have a good knowledge (maybe gpt-4o-mini) to assess the extracted qa and see if they are good to be asked in an interview or not