- https://arxiv.org/pdf/2312.10997.pdf
- Source: arxiv
TL;DR
A comprehensive survey that categorizes RAG systems into three paradigms — Naive, Advanced, and Modular RAG — and maps the full design space from indexing through evaluation.
Summary
Retrieval-augmented generation systems suffer from fragmented design decisions across indexing, retrieval, augmentation, and generation. This survey imposes structure by defining three evolutionary paradigms: Naive RAG, Advanced RAG, and Modular RAG. It covers sparse and dense retrieval, hybrid methods, chunking strategies, and augmentation techniques, then surveys evaluation frameworks including RAGAS and ARES. Persistent challenges such as hallucination, multi-hop reasoning degradation, and latency are analyzed alongside downstream applications in QA, code generation, and summarization.
Key contributions
- Proposes a three-paradigm taxonomy (Naive, Advanced, Modular RAG) that structures the entire RAG design space.
- Introduces a systematic review of indexing strategies — chunking, metadata, hierarchical — as a first-class design axis.
- Establishes hybrid retrieval (BM25 + bi-encoder) as consistently superior to either sparse or dense alone.
- Surveys automated evaluation frameworks RAGAS and ARES for reference-free assessment of RAG pipelines.
- Identifies multi-hop reasoning as the primary open challenge where current pipelines degrade with reasoning depth.
When to cite
- When framing a RAG system design and needing a taxonomy to position architectural choices.
- When justifying the use of hybrid retrieval over purely sparse or dense methods.
- When selecting an automated evaluation framework for a RAG pipeline without human annotation.
- When arguing that retrieval quality, not generation, is the primary source of hallucination in RAG systems.
LLM RAG Generative AI Prompt Engineering ARAGOG - Advanced RAG Output Grading RAG-Fusion - a New Take on Retrieval-Augmented Generation