Evidence completeness and answer quality in graph-assisted retrieval-augmented generation

Abstract

For multi-hop questions it remains unclear whether retrieving more complete evidence actually improves the answers of retrieval-augmented generation. The study compares sparse, dense and entity-mention-graph retrieval with a shared reranker and a local quantized language model on reserved HotpotQA questions, and proposes an evidence-to-answer decomposition that separates the effect of evidence coverage from the effect of evidence utilization. Graph-assisted retrieval kept complete evidence more often than dense retrieval, but its answer gain was not statistically resolved and was explained mainly by how the generator used the evidence rather than by retrieval coverage.

Keywords: retrieval-augmented generation, multi-hop question answering, entity mention graph, evidence coverage, answer quality

How to cite

Mukhambetov Bektas, Noken Rakhima, Tarlan Baimyrza, Yskak Zhanibek Evidence completeness and answer quality in graph-assisted retrieval-augmented generation // Proceedings of the International Scientific and Practical Conference “Science and Education in Central Asia: Digital Transformation and Sustainable Development” (CASJ-2026). — 2026. — P. 21–29. — URL: https://cajournal.kz/conf-paper.php?id=5

All conference papers