A Review of Retrieval-Augmented Generation in Natural Language Processing: Architectures, Challenges, and Future Research Directions

Abstract

Pre-trained Large Language Models (LLM) are known to have a number of parameters which implicitly contain world knowledge, and perform well on a wide range of Natural Language Processing (NLP) tasks. But these parameters are not transparent, not dynamic, not cheap to update, and can produce fluent but incorrect text, a phenomenon termed hallucination. Retrieval-Augmented Generation (RAG) bypasses these constraints, using a non-parametric memory, which is a collection of large text bodies, to be retrieved at inference time so that generation can be guided by retrieved evidence as well as by model weights. In this article we review the underpinnings of retrieval-augmented text generation as they evolved in the information retrieval (IR) and NLP literature up to 2021. With a retrieval-centric view, we break down the RAG paradigm into four steps: pre-retrieval, retrieval, post-retrieval, and generation, and explore key techniques for each step, such as approximate nearest neighbour indexing, sparse and dense retrieval, query modification, re-ranking and filtering, and generation conditioned by evidence. We present a taxonomy on representative systems, and analyse retrieval strategies along with algorithmic structure with a description of basic RAG workflow. We present an evaluation methodology review, comparative and meta-analytic tables summarising the field, and discussion of open challenges, retrieval quality, system efficiency, the role and selection of retrieval models, and promising directions for future research. The review brings a systematic categorisation and structure to early RAG studies, as well as a clear understanding of the underlying technology.