What a PRISMA Review Uncovers About Retrieval‑Augmented Generation (RAG)

1. Research Method: PRISMA 2020 Flowchart

Following PRISMA 2020 guidelines, the authors identified 4,721 records and after screening retained 128 high‑impact papers for systematic analysis.

Figure 1: Literature screening flow – 4,721 records identified, 128 papers included.

2. Technical Panorama of Retrieval‑Augmented Generation (RAG)

The surveyed works are organized into progressive stages, each introducing key innovations and representative approaches:

Pre‑retrieval : structure‑aware chunking (expanding from 100 to 4,000 tokens), metadata enrichment, long‑retrieval units – e.g., Chunking.

Retrieval : hybrid retrieval combining BM25, dense vectors, and knowledge graphs; graph traversal; dynamic triggering – e.g., Hybrid Retrieval.

Post‑retrieval : re‑ranking, context compression, noise injection, token budgeting – e.g., Post Retrieval.

Iteration Control : reflective tokens such as FLARE, RIND, Self‑RAG – e.g., Self‑RAG.

Memory Enhancement : user‑level vector stores, dialogue cache, knowledge‑graph integration – e.g., Memory.

Multi‑Agent Systems : tool‑chain orchestration (RALLE, MEDRAG) and ReAct‑Chain – e.g., Agentic.

Efficiency Compression : token‑level representations (xRAG) and pipeline scheduling (PipeRAG) – e.g., Efficiency.

Multimodal Retrieval : joint image‑text retrieval (MuRAG, Wiki‑LLaVA) – e.g., Multimodal.

These stages illustrate the evolution from a simple retrieve‑then‑generate pipeline to a programmable, explainable, and auditable distributed system.

3. Evaluation Metrics

Four metric families are commonly used to assess RAG systems:

Retrieval : Recall@k, MAP@k, Hit@k – measure recall performance of the retrieval component.

Generation : Exact Match (EM), F1, BLEU, ROUGE, BERTScore – evaluate textual quality of generated answers.

Hallucination : Support, Hallucination Rate, RAGTruth – assess factual consistency.

Human Evaluation : correctness, relevance, user satisfaction – capture subjective user experience.

LLM‑as‑Judge : GPT‑4 scoring, G‑EVAL, SelfCheckGPT – scalable model‑based evaluation.

4. Representative Datasets

The review catalogues 343 datasets; a subset of frequently used resources is listed below:

Natural Questions (NQ) – 323 k samples, open‑domain QA, cited 27 times.

HotPotQA – 113 k samples, multi‑hop QA, cited 26 times.

Wikipedia – 6 M articles, general‑purpose corpus, cited 19 times.

MS MARCO – 1 M passages, retrieval + QA, cited 8 times.

StrategyQA – 2.8 k samples, implicit reasoning, cited 8 times.

These datasets span a wide range of domains and scales, providing a comprehensive data landscape for RAG research. https://arxiv.org/pdf/2508.06401.pdf A Systematic Literature Review of Retrieval‑Augmented Generation: Techniques, Metrics, and Challenges