Relevance gain
UniProtKB corpus: Faithfulness +0.1176 and Answer Relevancy +0.1358 vs OpenAI.
Whitepaper
Domain-adapted retrievers paired with calibrated reranking outperform general-purpose providers across finance, clinical, and protein corpora.
UniProtKB corpus: Faithfulness +0.1176 and Answer Relevancy +0.1358 vs OpenAI.
Catastrophic errors drop from 15/97 to 1/97 (UniProtKB) and 5/97 to 1/97 (Clinical).
Clinical corpus: Faithfulness +0.0575, Answer Relevancy +0.0696 vs OpenAI.
Finance corpus: Answer Relevancy +0.0575 vs OpenAI (Holm-adjusted significance).
Prefer a new tab? Open the whitepaper directly.