slowbench
RetrievalInferenceEvalsAgentsSeriesArchive

Reference

Benchmarks

Every measurement on this site, in one place. Each table links to the post that produced it, where the setup, the caveats and what was inferred rather than measured are written down.

Tables
1
Rows
4

Reranker comparison — Pass@k, MRR, wall clock

42 Korean issue-title queries · hybrid dense+BM25 index · top-50 candidates reranked

RerankerPass@1Pass@5Pass@10MRRWall clock
none26.2%45.2%47.6%0.33333s
jina-reranker-v1-turbo-en23.8%42.9%47.6%0.32654s
jina-reranker-v2-multilingual26.2%42.9%47.6%0.330282s
bge-reranker-v2-m328.6%42.9%47.6%0.352514s

Pass@10 is identical in all four runs. Reranking reorders a candidate set; it cannot add to it.

Source · Reranking cannot find what retrieval missed