GAP: Graph-Anchor Pyramid for Resolving Relational Blindspots and Recursive Semantic Decay in Hierarchical RAG

While hierarchical aggregation algorithms successfully scale the prompt token window of Large Language Models (LLMs), flat text-based reductions suffer from two fundamental challenges: a "Multi-Hop Relational Blindspot" in retrieval and "Recursive Semantic Decay" during synthesis. Standard similarity searches fail to retrieve disjoint root causes that do not share explicit lexical features with the query, while the lossy nature of natural language tree-reductions aggressively prunes granular entity-relation links before they reach the final root summary. We propose GAP (Graph-Anchor Pyramid), a unified hierarchical ingestion and aggregation framework. By capturing causal dependencies during retrieval, grouping connected nodes topologically, and injecting semantic anchors in prompts, GAP resolved the multi-hop relational blindspot and recursive semantic decay, achieving 100.0% causal recall under tight length constraints across model generations.