Nugget-Level Retrieval Coverage Prediction for Complete and Faithful RAG Answers

Abstract

Retrieval-augmented generation (RAG) depends on retrieved passages to provide the evidence from which an answer is composed. Conventional retrieval metrics reward relevance, yet a highly relevant top-k list can repeatedly surface the same fact while omitting other information units required for a complete response. This study investigates nugget-level retrieval coverage prediction on CoverageBench, which comprises seven retrieval collections, 334 topics, and six retrieval configurations. Using the benchmark’s aggregate nDCG@20, alpha-nDCG@20, and Subtopic Recall at rank 20 (StRecall@20) measurements, we introduce NLCP-Ridge, a deterministic ridge-regression model that predicts coverage from relevance, diversity-aware relevance, and reranker identity. Leave-one-dataset-out validation over 42 held-out system-dataset cases yields MAE = 0.0464, RMSE = 0.0687, R2 = 0.9090, Pearson r = 0.9628, and Spearman rho = 0.9713. Relative to an nDCG-only predictor, NLCP-Ridge reduces MAE by 58.6%; relative to an alpha-nDCG-only predictor, it reduces MAE by 18.0%. The retrieval comparisons further show that the highest-nDCG configuration is not always the highest-coverage configuration: CAsT and CRUX-MultiNews select different winners under StRecall@20. The findings show that a compact coverage estimate can complement relevance evaluation and provide an evidence-sufficiency signal before answer generation.