AFRIGROUND: Context-Aware Retrieval and Grounded Question Answering for African Population-Health Knowledge

Retrieval-augmented systems increasingly mediate access to health information, yet semantic relevance alone does not ensure that retrieved evidence is appropriate to a population, place, time, or health indicator. We introduce AFRIGROUND, a context-aware retrieval and knowledge exploration framework for evidence on African population health. Built from Demographic and Health Survey (DHS) resources, AFRIGROUND contains 309,188 retrieval-ready knowledge points with preserved geographic, temporal, indicator, and One Health metadata. We evaluate BM25, Multilingual E5, BGE-M3, and Reciprocal Rank Fusion (RRF) on a 60-query benchmark with 3,343 blinded graded relevance judgments. Multilingual E5 achieves the highest aggregate performance (P@5=0.893; nDCG@10=0.886; Indicator Match@10=0.890), while BGE-M3 performs statistically comparably; both dense retrievers and Hybrid RRF significantly outperform BM25. AFRIGROUND further translates the evaluated retrieval framework into an interactive dashboard for grounded conversational question answering, evidence inspection, country and One Health exploration, and retrieval-model comparison.