Main comparison

Results

All six models, evaluated identically against the same real NFCorpus test split (n=323 queries) and the same qrels. No number below is estimated — every cell is generated from results/metrics/*.json.

Model	P@10	Recall@100	MAP	MRR@10	nDCG@10	Latency (ms/query)
TF-IDF	0.2167	0.2372	0.1372	0.5062	0.3050	3.631
BM25	0.2071	0.2295	0.1333	0.4939	0.2954	2.155
BGE	0.2796	0.3368	0.1831	0.5556	0.3712	2.678
MedCPT	0.2697	0.3488	0.1824	0.5487	0.3654	1.871
BM25+MedCPT (RRF)	0.2598	0.3389	0.1809	0.5678	0.3620	4.069
Hybrid+Reranker (pool=50)	0.2765	0.2782	0.1760	0.5670	0.3731	1945.6

Bold/accent = best value in that column. Reranked row's Recall@100/MAP are capped by its 50-document candidate pool — see the note below and Evaluation for the full explanation and statistical significance tests.

Reading the reranker row correctly

A cross-encoder can only reorder the candidates it is given. With a candidate pool of 50, the reranked run never contains more than 50 documents per query, so Recall@100 == Recall@50 exactly by construction — confirmed mechanistically: at pool=100, Recall@100 exactly matches hybrid RRF's own Recall@100 (reordering an unchanged 100-document set cannot change how many relevant documents it contains). This is a genuine methodological property, not a bug — see the pool-size ablation on Efficiency.

Progression across the study

BM25<TF-IDF<Hybrid RRF<MedCPT<BGE<Hybrid+Reranker

Ordered by real nDCG@10, lowest to highest.