Eighty years ago, long before the first digital computer had earned its place in history, Vannevar Bush imagined the “memex,” a machine that would extend human memory by letting us retrieve, on demand, the exact piece of knowledge a moment required. Half a century of information retrieval followed, from the inverted index and the vector space model to the search engines that now organize the web. Today, large language models (LLMs) such as GPT-5 have astounded us with their ability to generate coherent, contextually relevant text, seemingly bridging the gap between human and machine communication. Yet, for all their fluency, these models reason only over the frozen statistical patterns absorbed during training: they cannot tell us where an answer came from, cannot be updated without being retrained, and will, with great confidence, invent facts that were never true. “Understanding retrieval-augmented generation,” i.e., understanding how a machine can ground what it generates in external knowledge it can search, cite, and inspect, is a significant milestone in our journey toward AI we can actually trust. 

This textbook delves into the contract at the heart of that idea: a fluent but forgetful generator paired with a retriever that supplies it, at the moment of answering, with the most relevant evidence from an external memory of documents. LLMs excel at producing text that reads as intelligent and meaningful, but their limitations become evident whenever a task demands current, verifiable, or domain-specific knowledge. Retrieval-augmented generation (RAG) addresses this by separating what a model knows from what it can look up, so that answers are anchored in sources a reader can examine rather than conjured from parameters alone. The author traces the full lineage of the field, from Boolean and ranked retrieval, term weighting, and link analysis to dense and hybrid retrieval, re-ranking, and the modern RAG pipeline, showing how half a century of classical ideas quietly underwrites today’s most capable systems.

The textbook explores the current state of RAG, its capabilities and limitations, and the technical foundations required to build systems that are not only fluent but grounded: representation and indexing, sparse and dense retrieval, ranking and evaluation, faithfulness and honest measurement, agentic RAG, and the pursuit of efficient, neurosymbolic intelligence. It also examines the ethical implications and societal impacts of these systems, from misinformation and transparency to the energy cost of ever-larger models. Containing exercises, assignments, and a comprehensive quiz, and complemented by an integrated survey of fundamental RAG algorithms, the textbook is meant as a reference for courses on information retrieval, AI, NLP, data analytics, data mining, and more.

Erik Cambria is a Professor at Nanyang Technological University, where he also holds the appointment of Provost Chair in Computer Science and Engineering, and Founder of several AI companies, such as SenticNet, offering B2B sentiment analysis services, and finaXai, providing fully explainable financial insights. Prior to moving to Singapore, he worked at Microsoft Research Asia (Beijing) and HP Labs India (Bangalore), after earning his PhD through a joint program between the University of Stirling (UK) and MIT Media Lab (USA). Today, his research focuses on neurosymbolic AI for interpretable and explainable affective computing in domains like mental health, climate resilience, and socially responsible investing. He is ranked in Clarivate’s Highly Cited Researchers List of World’s Top 1% Scientists, is recipient of many awards, e.g., IEEE Outstanding Early Career, was listed among the AI’s 10 to Watch, and was featured in Forbes as one of the 5 People Building Our AI Future. He is an IEEE Fellow, Associate Editor of various top-tier AI journals, e.g., Information Fusion and IEEE Transactions on Affective Computing, and is involved in several international conferences as keynote speaker, program chair and committee member.