Open access
Jul 2026
ADAPTIVE MULTI-STAGE VECTOR RETRIEVAL FOR RETRIEVAL-AUGMENTED GENERATION
The Adaptive Multi-Stage Vector Retrieval (AMSVR) framework is proposed, prioritising weighted, drift-resistant composition over uniform fusion, and offers tailored configurations: AMSVR-Scientific (dense + tuned hybrid) peaks at NDCG@10 = 0.7570 on SciFact, while AMSVR-Full (seven stages) targets broader, noisier corpora where Recall@100 matters most.
Samsudeen Alabi Bankole, Yakub Kayode Saheed
· NLP & Big Data · 0 citations