An Improved Retrieval-Augmented Generation Agent for UAV Situation Awareness Report Generation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Accurate and timely UAV (Unmanned Aerial Vehicle) situation awareness is essential for early warning and operational decision-making in low-altitude security, yet the manual analysis of large-scale heterogeneous data is resource-intensive and introduces delays. This study proposes IRAG-UAVSA (Improved Retrieval-Augmented Generation—Unmanned Aerial Vehicle Situation Awareness) Agent, an autonomous situation-awareness report-generation framework based on an improved retrieval-augmented generation architecture. The framework integrates real-time UAV descriptors, including platform type, intent, and threat level, with four task-specific hierarchical knowledge bases. A dual-loop pipeline combines evidence-template-guided hybrid retrieval and cross-encoder reranking with high-level LLM (Large Language Model) reflection, structured evidence filling, rule-based validation, and threshold-triggered iterative retry. A three-level evaluation framework was designed, comprising process-level RAG metrics, automatic natural language evaluation, and human–LLM collaborative scoring. Across 49 domain-screened test inputs, IRAG-UAVSA Agent achieved 93.0% knowledge coverage, 4.1% hallucination, and 91.7% logical consistency, while also obtaining the best SummaC (0.67) and VeriScore (0.83) results among the evaluated baselines. The results indicate that iterative retrieval with explicit evidence constraints can substantially improve the factual grounding, traceability, and practical utility of LLM-generated UAV situation-awareness reports.