Why Classical Information Retrieval Is the Missing Foundation of Modern RAG Systems

Many engineers building Retrieval-Augmented Generation (RAG) systems struggle with high costs and fragile architectures because they treat RAG as a new invention rather than a classical search problem with a large language model attached. The article argues that studying Information Retrieval (IR), particularly the textbook by Manning, Raghavan, and Schütze, offers proven solutions to these structural issues. Lexical search algorithms like BM25 complement dense vector embeddings by handling exact keyword matches — such as product codes and error IDs — without relying on neural networks that can become outdated. A multi-stage retrieval pipeline, where cheap methods like BM25 first narrow results to a candidate set before an expensive re-ranker refines the top results, can dramatically cut infrastructure costs. IR also provides rigorous evaluation metrics such as Precision@K and Recall, replacing guesswork with measurable proof that a retrieval system is actually working.