AURORA-RAG: A Trustworthy Hierarchical Retrieval-Augmented Generation Architecture for Grounded Synthetic Personas

Large language models can generate synthetic personas, but their factual claims may be unsupported when generation relies on parametric knowledge alone. We present AURORA-RAG, a hierarchical retrieval-augmented generation architecture that introduces documentary evidence at country–topic, community, community–topic, and persona construction stages. The evaluated four-country corpus comprises 834 input files and 3888 indexed semantic chunks, with non-English source material translated into English. AURORA-RAG combines dense retrieval and BM25 using reciprocal rank fusion and pseudo-relevance feedback while retaining document provenance. We compare no-RAG, persona-only hybrid RAG, and full AURORA-RAG using 1440 personas per condition across three geopolitical topics. Corpus support increases from 68% to 82% and 91%, respectively, while unsupported-claim rates decline from 24% to 13% and 7%. Between retrieval-enabled conditions, supplied-evidence faithfulness rises from 79% to 90%, attribution accuracy from 82% to 92%, and citation correctness from 85% to 94%. Full AURORA-RAG also produces greater opinion displacement from no-RAG than persona-only retrieval (0.210 versus 0.103), a measure of output change rather than accuracy. Hybrid retrieval leads on five of six ranking metrics, although only MRR@5 over dense retrieval remains significant after Holm correction. These findings support hierarchical retrieval for stronger documentary grounding and source traceability.