Retrieval Metrics Demystified: From BM25 Baselines to EM@5 & Answer F1
What BM25, Recall@k, answer-in-context rate and token F1 actually measure, which number to fix first, and how to compute them without fooling yourself.
Retrieval and RAG01
Series · 2 parts · 2,024–2,025
Which retrieval numbers actually matter, and a reproducible way to find out which stack to build on.
What BM25, Recall@k, answer-in-context rate and token F1 actually measure, which number to fix first, and how to compute them without fooling yourself.
Retrieval and RAG01
How to benchmark sparse, dense, hybrid and late-interaction retrieval on your own corpus, and how to read the numbers without fooling yourself.
Retrieval and RAG02