Answers are grounded in this paper and cite their sources.
Large language models (LLMs) suffer from two persistent problems in their parametric knowledge: update lag and factual deviation, which hinder their deployment in professional domains. Retrieval-augmented generation (RAG) connects external corpora at inference time, expanding the knowledge boundary without modifying model parameters, and has become the mainstream solution for enterprise knowledge bases, intelligent question answering, and document understanding. This paper divides RAG technology into four stages along its design focus: naive retrieval, precise retrieval, adaptive routing, and agentic reasoning, and examines the principal contradictions and new costs of each stage. On this basis, a hierarchical evolution framework is proposed that cross-compares six levels—retrieval, knowledge, fusion, generation, conflict arbitration, and engineering evaluation—against the four stages, identifying gaps in existing research, especially the unresolved arbitration of multi-source contradictory documents across all stages. The paper also discusses applicable boundaries: when knowledge is already well mastered by the model, when the context window suffices to cover the full corpus, or when the knowledge base is small and sparsely updated, adding a retrieval link only increases latency and noise. Finally, selection guidance is provided along three dimensions: accuracy, computational cost, and robustness, serving as a reference for engineering design and future research.
No takes yet. Share an insight, caveat, or question.
Kaijie Zhuang (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: