Hi, I was trying to determine what the optimal Top k values are and I experimented with values ranging from (5-10). I noticed that hit rate increases and MRR remains quite similar when I use a higher Top K value . Does this mean that using a higher Top K value is better or do i need to consider other factors as well ?
One factor to consider is that LLM synthesis tends to degrade in quality when dealing with large amounts of context (it starts missing stuff easier). Also increased cost & latency