Retrieval-Augmented Generation for Large Language Models: Architectures, Retrieval Optimization, Evaluation, Robustness, and Adaptive Evidence-Centric Design

Retrieval-Augmented Generation (RAG) has become a central paradigm for improving the factuality, freshness, and domain adaptability of large language models (LLMs) by conditioning generation on externally retrieved evidence. The supplied seminar report provides a broad foundation covering LLM limitations, RAG architecture, retrieval strategies, augmentation mechanisms, RAG paradigms, applications, evaluation, and open challenges. This article restructures that material into a journal-oriented review and extends it with a research taxonomy centered on retrieval quality, evidence integration, generation faithfulness, robustness, and computational efficiency. Recent literature indicates that RAG evaluation must move beyond isolated retrieval or generation metrics toward end-to-end, multi-dimensional assessment, while contemporary systems increasingly employ hybrid retrieval, reranking, adaptive retrieval, modular pipelines, and domain-specific optimization. Based on the identified limitations, this review proposes an Evidence-Centric Adaptive RAG (ECAR) framework in which retrieval decisions, evidence quality, context compression, generation constraints, and post-generation verification are treated as coordinated components rather than independent modules. The framework is conceptual rather than experimentally validated and is intended to define a testable research agenda. The review further develops a comparative taxonomy, identifies persistent research gaps, and proposes an experimental protocol based on retrieval quality, answer correctness, faithfulness, robustness, latency, and cost. The resulting manuscript provides a stronger foundation for a publishable review article and for a subsequent empirical dissertation study.

کلیدواژه ها:

Retrieval-Augmented Generation ، Large Language Models ، Information Retrieval ، Hybrid Retrieval ، Adaptive Retrieval ، Hallucination ، Faithfulness ، Reranking ، RAG Evaluation ، LLMs