Abstract
<title>Abstract</title> <p>Traditional retrieval-augmented generation pipelines partition target corpora into isolated text segments, omitting structural document hierarchies and failing to resolve relational hyperlink topologies. Consequently, standard single-pass search architectures struggle with boundary fragmentation, vocabulary mismatch, and high semantic drift. To resolve these limitations, this paper introduces a high-performance hybrid retrieval engine that unifies dense semantic search, sparse lexical search (BM25), classical TF-IDF search, and a native distributed file-system retrieval layer. The architecture implements an over-fetching candidate protocol coupled with customizable hybrid weighting and configurable combination modes (Sum, Max, Mean) to address the scale mismatch between divergent scoring spaces. Furthermore, a topologically-aware multi-tier retrieval controller is developed to automatically parse primary candidate documents, identify hyperlinked references to external collaborative enterprise document nodes, and execute a deterministic secondary retrieval pass to capture supporting context. By utilizing local context-aware cosine reranking over a global cycle-avoidance traversal graph, the system preserves structural semantic continuity without incurring the high computational overhead and latency associated with language-model-driven query decomposition. Systematic evaluations demonstrate that the unified pipeline secures up to a +20.0% improvement in absolute recall (reaching 1.0000 Recall@100 on multi-hop benchmarks), a +13.2% gain in ranking quality (NDCG@10 increasing from 0.7654 to 0.8664), and a >95% reduction in context-processing latencies (replacing multi-second LLM query decomposition loops with sub-millisecond 0.87 ms topological anchor reranking), establishing a scalable paradigm for real-time enterprise information extraction. We evaluated the system using an open-source benchmark harness. All experiments, datasets, configurations, and raw outputs are available in the accompanying GitHub repository.</p>