medium10 pts · first answer countscontext-windowrag

Facts buried mid-context

With 40 retrieved chunks in the prompt, the model often misses facts placed in the middle, though it finds them at the start or end. What helps?

A.Add more chunks so the middle is biggerB.Send fewer, better-ranked chunks and put the most relevant ones at the start or end near the questionC.Remove the question from the promptD.Turn all chunks into one long paragraph

← Every reply sounds like example 1 Parsing an answer after reasoning →