From my understanding, a typical "AI" Q&A system has a (vector) database of embedded text (from a set of documents). And when a user asks a question, the user's question is embedded and some similarity metric (cosine similarity) is used to extract the top k relevant embeddings from the vector database pertaining to the question. Then, usually most relevant embedding is used to summarize an answer back to the user.
My question is would it be possible if a better answer exists that is some combination of the top K relevant embeddings. And if so, what would be the most efficient way to extract that better combination? The only idea I can come up with is trying every combination of top K embeddings (i.e. for K=3 ==> (1), (2), (3), (1,2) (1,3), (2,3), (1,2,3)), but this idea seems computationally expensive.
My inspiration for this question came from the fact that when you are writing a research paper or studying, you usually try to synthesize a singular answer from multiple source. It's not usually the case that you get the top 3 (K) results and pick the top one that is most relevant.
I know it's a long post, but I really couldn't find an answer to this type of question anywhere. Also, I can't tell if it is not a relevant question for some scientific, trivial reason. I'm also new to NLP, so apologize if the answer to this question is trivial.
I would love for some answers regarding this or some direction as to where I can learn more!