Retrieval-Augmented Generation (RAG) has emerged as a transformative paradigm for enhancing Large Language Models (LLMs) by grounding responses with external, non-parametric knowledge to enhance outputs and reduce hallucinations; however, the naïve retrieve-then-read approach faces significant limitations, necessitating optimization across the entire pipeline. This comprehensive survey delves into the diverse landscape of these optimization methodologies, moving beyond traditional approaches to explore innovative strategies that address critical challenges. We propose a four-stage taxonomy (Indexing, Retrieval, Augmentation, Generation) to structure the landscape of RAG optimization techniques, providing a detailed examination of chunking strategies, embedding techniques, query expansion methods, context compression algorithms, and prompt engineering approaches. Furthermore, we investigate recent innovations such as adaptive retrieval, hybrid architectures, and joint optimization techniques that synergistically combine improvements across multiple stages for enhanced end-to-end performance. Based on our synthesis of the existing literature, significant challenges remain in the development of effective RAG systems. Specifically, further research is needed to establish robust evaluation metrics, address computational limitations, and adapt these systems for use in diverse domains.
Paper
Full text
Towards Optimized Retrieval-Augmented Generation: A Comprehensive Survey
Semantic Scholar · 2025
Abstract
Retrieval-Augmented Generation (RAG) has emerged as a transformative paradigm for enhancing Large Language Models (LLMs) by grounding responses with external, non-parametric knowledge to enhance outputs and reduce hallucinations; however, the naïve retrieve-then-read approach faces significant limitations, necessitating optimization across the entire pipeline. This comprehensive survey delves into the diverse landscape of these optimization methodologies, moving beyond traditional approaches to explore innovative strategies that address critical challenges. We propose a four-stage taxonomy (Indexing, Retrieval, Augmentation, Generation) to structure the landscape of RAG optimization techniques, providing a detailed examination of chunking strategies, embedding techniques, query expansion methods, context compression algorithms, and prompt engineering approaches. Furthermore, we investigate recent innovations such as adaptive retrieval, hybrid architectures, and joint optimization techniques that synergistically combine improvements across multiple stages for enhanced end-to-end performance. Based on our synthesis of the existing literature, significant challenges remain in the development of effective RAG systems. Specifically, further research is needed to establish robust evaluation metrics, address computational limitations, and adapt these systems for use in diverse domains.