Exploring Generative Retrieval: Memory Mechanisms, GDR Paradigm, and Practical Applications

In this talk, the speaker introduces the topic of generative retrieval (GDR), reviewing common retrieval paradigms—sparse (TF‑IDF, BM25), dense (DSSM, contrastive learning), and generative approaches—and presents an EACL 2024 oral paper that analyzes the memory‑mechanism challenges of GDR.

The research team from Xiaohongshu, consisting of four members focusing on large‑model evaluation, inference, and generative retrieval, outlines the GDR pipeline: a query encoder produces a vector, which is fed to a decoder that autoregressively generates note IDs, leveraging the model’s parameters as a memory store.

Experiments compare GDR with dense retrieval (AR2) and a baseline generative model (NCI), showing GDR’s superior recall on small candidate sets but performance degradation as the candidate pool grows, limited memory capacity, and difficulties updating the index when documents change.

To mitigate these issues, the authors propose a hybrid architecture: use GDR’s memory to retrieve high‑level document clusters, then apply dense matching within the selected clusters for fine‑grained relevance, along with adaptive negative sampling and memory‑friendly cluster construction.

Application scenarios include online streaming note recall, high‑frequency query optimization, full‑process query suggestion, and future extensions such as integrating LLMs into GDR and adaptive efficiency‑aware retrieval strategies.

The session concludes with a Q&A addressing hierarchical clustering stability, handling new documents, multi‑cluster assignments, and real‑world deployment progress.