A Systematic Review of Key Retrieval-Augmented Generation (RAG) Systems: Progress, Gaps, and Future Directions

Oche et al. 2025 — RAG Systematic Review

Citation

Oche, A. J., Folashade, A. G., Ghosal, T., & Biswas, A. (2025). A Systematic Review of Key Retrieval-Augmented Generation (RAG) Systems: Progress, Gaps, and Future Directions. arXiv preprint arXiv:2507.18910.

Framing

This is a comprehensive systematic review that traces RAG’s evolution from 2017 through mid-2025. The unique contribution is the year-by-year progress mapping combined with analysis of enterprise deployment challenges, particularly proprietary data handling and emerging agentic architectures.

The authors position RAG as the solution to a fundamental LLM limitation: models store knowledge in parameters but can’t easily update that knowledge or verify its accuracy. RAG introduces “non-parametric memory” via retrieval, creating a hybrid system that combines the generative fluency of LLMs with the verifiability and currency of retrieved documents.

Core Argument

RAG represents a shift from purely parametric AI (all knowledge in model weights) to hybrid memory systems that combine parametric knowledge with retrieval-based external knowledge. This enables:

1. Factual grounding: outputs can be traced to source documents
2. Knowledge currency: can update without retraining
3. Hallucination reduction: generation is conditioned on retrieved evidence
4. Scale efficiency: retrieval can substitute for model size

Key Milestones Documented

• 2017-2019: Pre-RAG retrieve-and-read pipelines (DrQA, ORQA)
• 2020: RAG formally introduced (Lewis et al.); DPR; REALM
• 2021: Fusion-in-Decoder; KILT benchmark; expansion beyond QA
• 2022: RETRO (retrieval substitutes for scale); Atlas (few-shot)
• 2023-2024: Integration with larger LLMs; enterprise deployment
• 2025: Agentic RAG architectures; privacy-preserving techniques

Atoms Extracted

• 05-atom—rag-definition
• 05-atom—retrieval-substitutes-for-scale
• 05-atom—parametric-vs-non-parametric-memory
• 05-atom—rag-fusion-strategies
• 05-atom—rag-hallucination-reduction-mechanism

Molecules Extracted

• 05-molecule—rag-technical-architecture
• 05-molecule—rag-enterprise-challenges

Notes

The paper covers 2017–mid-2025, making it one of the most current comprehensive RAG reviews available. Particularly valuable for understanding the trajectory from simple retrieve-read pipelines to sophisticated agentic systems. The enterprise focus (Sections 5 and 7) is especially relevant for practical implementation questions.

Related:, 05-atom—agentic-ai