The rapid adoption of large language models (LLMs) across many domains has highlighted Retrieval‑Augmented Generation (RAG) as a vital approach to curb hallucinations and boost performance. Numerous RAG methods have been introduced, each with distinct features, making the selection of a suitable technique for a given scenario challenging. This work proposes a conceptual framework that systematically characterises (what it is) and analyses (how well it does) existing RAG approaches. The framework defines five primary dimensions: retrieval source (internal knowledge base, external web, hybrid), retrieval method (dense vectors, sparse inverted index, hybrid), fusion strategy (pre‑retrieval, post‑retrieval, cross‑attention), generation control (prompt engineering, fine‑tuning, reward modeling), and evaluation metrics (accuracy, robustness, cost). Each dimension contains sub‑dimensions that capture implementation details and performance aspects. The framework was derived from a qualitative analysis of attributes reported in the literature and serves as a pre‑empirical filter, enabling researchers and practitioners to narrow down RAG candidates that satisfy specific user and task requirements before costly benchmarking.
Its main contribution is a structured language for describing both the nature and effectiveness of RAG techniques, promoting transparent and systematic comparison and laying the groundwork for more efficient empirical studies.
Review