Retrieval-Augmented Generation (RAG) Architectures: GraphRAG, Hybrid Reranking, and Context Drift Mitigation

Abstract—Retrieval-Augmented Generation (RAG) has emerged as the definitive architectural paradigm for grounding Large Language Models (LLMs) in private enterprise corpora, mitigating hallucinations, and dynamically expanding knowledge without retraining. However, conventional 'Naive RAG' systems rely on flat vector-similarity search over isolated text chunks, exposing profound structural vulnerabilities: semantic fragmentation, complete inability to execute multi-hop global summarization, and acute susceptibility to the 'lost-in-the-middle' context degradation effect. To address these systemic failure modes, modern information systems are converging toward Advanced and Modular RAG architectures. This paper delivers an exhaustive, mathematically formal technical investigation and empirical benchmark of modern RAG paradigms. We formulate the probabilistic marginalization dynamics of retrieval-generation pipelines, model the structural topology of Graph-Augmented RAG (GraphRAG) with hierarchical community detection, and evaluate hybrid multi-stage retrieval combining dense dual-encoders, sparse lexical algorithms (BM25), Reciprocal Rank Fusion (RRF), and cross-encoder neural rerankers. Furthermore, we construct an enterprise benchmarking testbed across three complex document corpora (financial multi-10-K filings, biomedical clinical literature, and legal case law). Our empirical results demonstrate that while naive vector RAG suffers an accuracy collapse to 38.2% on multi-hop global sensemaking queries, GraphRAG achieves 84.6% answer comprehensiveness. Concurrently, hybrid multi-stage reranking reduces context hallucination by 58.4% while maintaining sub-second query latency. Finally, we formulate open research directions in dynamic context pruning, agentic iterative retrieval loops, and post-retrieval factual provenance verification. Index Terms—Retrieval-Augmented Generation (RAG), GraphRAG, Knowledge Graphs, Hybrid Search, Neural Reranking, Context Drift, Large Language Models (LLMs), Reciprocal Rank Fusion.