Announcement
CShorten30 reports reranking gains, but model fusion mostly runs up the bill
Combining three LLM rerankers costs 15 times GPT-5.6 Luna alone without beating GPT-5.6 Terra, CShorten30 reports.
TLDR
In “Scaling Reranked Retrieval,” CShorten30 reports that deeper retrieval plus zerank-2 raises nDCG@10 from hybrid retrieval’s 0.181 to 0.424; a GPT-5.6 Luna pass over the top 20 then reaches 0.470. Throwing more documents at that last stage brings diminishing returns: CShorten30 says expanding it to 100 documents yields 0.482 at five times the cost.
Combined views
53
20 Sources, first seen ago

Comments