AURORA-RAG: A Trustworthy Hierarchical Retrieval-Augmented Generation Architecture for Grounded Synthetic Personas
Year: 2026
Type: article
Abstract: Large language models can generate detailed synthetic personas, but unconstrained generation can produce unsupported or contextually inconsistent outputs. We introduce AURORA-RAG, a hierarchical Retrieval-Augmented Generation framework that incorporates documentary evidence at country–topic, community, community–topic, and persona stages. The system combines dense retrieval and BM25 through Reciprocal Rank Fusion, applies Pseudo-Relevance Feedback, preserves source provenance, and restricts final evidence to indexed corpus passages. We evaluate AURORA-RAG using 1,440 personas across four national environments and three geopolitical topics. Full hierarchical retrieval produces greater downstream conditioning than persona-only retrieval, with mean displacement from no-RAG increasing from .103 to .210. Corpus-supported claims increase from 68% under no-RAG to 82% under single-stage RAG and 91% under full AURORA-RAG, while hallucination frequency decreases from 24% to 13% and 7%. Evidence faithfulness rises from 79% to 90%, attribution accuracy from 82% to 92%, and citation correctness from 85% to 94%. Hybrid RRF achieves the highest observed value on five of six retrieval metrics. These results support hierarchical retrieval as a trustworthy RAG design for improving documentary grounding, reducing unsupported generation, and increasing the auditability of LLM-based synthetic systems.
Source: Electronics
Authors: Rebecca Marigliano, Kathleen M. Carley
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,896
OpenAlex ID: W7213642249