Recall@10 reportedly peaked around 100 documents, then fell, in a cross-encoder reranking test
Weaviate Podcast highlights Mathew Jacob's experiments with reranking retrieved documents and says another approach—sliding-window listwise reranking—stayed robust at 1,000 documents.
TLDR
Weaviate Podcast says Mathew Jacob tested increasing the number of documents a cross-encoder reranks while interning at Databricks. Recall@10 climbed, peaked around 100 documents and then declined. According to the podcast, Jacob checked for a bug and found none. Episode #141 features Jacob, lead author of “Drowning in Documents,” and explores reranking depth, sliding-window listwise reranking that it says stayed robust at 1,000 documents, and ranking cascades. It also discusses TraceLab, described as 40,000 real traces of coding agents at work.
Combined views
—
1 Source, first seen ago
