[bardiapour]

i think not, better results >>> better latency + cost

[robrenaud]

Maybe a combined approach beats either? Let some non-LLM reranker quickly spit out two results, and fill in the rest with the LLM.