Recall@10 reportedly peaked around 100 documents in a reranking experiment

Weaviate Podcast says Mathew Jacob tested how many retrieved documents a cross-encoder should rerank while interning at Databricks. Recall@10—recall among the top 10 results—rose, peaked around 100 documents, then declined, the podcast says. Episode 141 features Jacob, lead author of "Drowning in Documents," and explores sliding-window listwise reranking that the podcast describes as robust at 1,000 documents. It also discusses ranking cascades and TraceLab, described as 40,000 real traces of coding agents at work.