The Chunking Crisis in Hierarchical Structural Law: Bridging the Ontological Gap in Legal RAG via Deterministic Ontology-Driven Ingestion

Standard Retrieval-Augmented Generation (RAG) architectures struggle with hierarchical legislation because chunking separates listed entities, such as protected species in annexes, from their governing articles. To bridge this ontological gap, we propose Semantic Breadcrumbs, a deterministic retrieval framework whose ontology-driven ingestion prefixes each chunk with its structural hierarchy and legal category. At inference time, a two-hop retriever anchors the questioned entity under a phrase-level identity gate, fetches the governing rules through citation-priority indexing, and routes a lightweight generative model into route-specific prompt contracts ending in a typed legal status. We evaluated the architecture on maritime and environmental legislation from nine Francophone African jurisdictions, using 104 citation-verified list-to-rule questions, control arms, and exact paired tests with Holm correction. Semantic Breadcrumbs reached 91.35% accuracy and a 91.43% macro F1-score, against 55.77% and 54.59% for standard chunking, a significant difference after correction. It was more accurate than exhaustive annex expansion with 2.6 times less context, a difference that is not significant after correction. Four context-aware chunking baselines at a matched budget trail it significantly after correction. It converted 86.21% of hallucination traps into typed abstentions. The evidence covers list-to-rule questions with one embedding model; a further generator replicates the ordering.