The insights on "few-shot" learnings are very interesting. We also wanted to use a vector db combined with GPT for our search engine. Intuitively, I would have thought the accuracy with this method to be much higher. Models hallucinating or giving wrong answers +20% of the time is not possible for direct answers, as helpers it's a good idea!

Wondering if there would be a way to create a score of confidence to the answers from this method? And discard the lowest ones.