Robust-GAP: Hallucination-Resistant Hierarchical RAG with Dynamic Topology and Citation Provenance

Retrieval-Augmented Generation (RAG) has transformed enterprise document analysis. However, state-of-the-art hierarchical synthesis methods, such as the Graph-Anchor Pyramid (GAP), suffer from three critical bottlenecks in mission-critical applications (e.g., Site Reliability Engineering [SRE]): (1) high maintenance overhead of static topology, (2) lack of source tracking (citation provenance) during hierarchical reduction, and (3) cascading hallucination propagation. In this paper, we propose Robust-GAP, an autonomous hierarchical RAG framework designed for high-noise operational environments. Robust-GAP introduces Dynamic Lightweight Causal Extraction (DLCE) to construct directed local causal graphs on-the-fly from unstructured metadata. It employs Schema-Guided Anchor Validation (SGAV) to filter hallucinatory or irrelevant entity-relation triplets, preventing error propagation. Furthermore, it incorporates Provenance-Aware Pyramid Propagation (PAPP) to track raw document IDs hierarchically, enforcing inline citations [Dxx] in final outputs. We evaluate Robust-GAP using two Large Language Models (LLMs) on a 32-node enterprise incident log dataset. Experimental results demonstrate that Robust-GAP achieves a 100% causal chain recall and 100% citation F1-score with 0.0% hallucination rate under severe noise, significantly outperforming conventional one-shot RAG and classic GAP baselines.