Generative Retrieval

Generative retrieval (GR) is a novel information retrieval paradigm that uses generative language models to directly produce identifiers of relevant documents or passages, bypassing traditional ranking methods. Current research focuses on improving the accuracy and efficiency of GR, exploring various identifier types (numeric, textual, multi-modal), and addressing challenges like hallucination and scalability through techniques such as few-shot indexing, self-consistent reranking, and the integration of dense retrieval methods. GR's potential impact lies in its ability to enhance efficiency and generalization in search, recommendation systems, and knowledge-intensive tasks, offering more flexible and potentially more human-interpretable results.