Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces ReDSI, an open-source implementation of the differentiable search index (DSI) framework that addresses reproducibility issues by supporting all three document identifier types: atomic, naive, and semantic. The authors provide a comprehensive NQ320K dataset construction pipeline and demonstrate that ReDSI achieves competitive or superior results compared to existing DSI baselines. Extensive experiments on model downscaling reveal insights into retrieval effectiveness, parameter efficiency, and training methods, paving the way for future research in generative retrieval.
ReDSI not only standardizes DSI implementations but also outperforms previous baselines, challenging the status quo in generative retrieval.
The differentiable search index (DSI) framework (Tay et al., 2022) has become the de facto baseline for generative retrieval. However, DSI is hard to reproduce: no public implementation covers all three original document identifier types (atomic, naive, semantic), reported results vary widely, and the ubiquitous NQ320K dataset is built from Natural Questions through diverse and underspecified preprocessing. We introduce ReDSI, the first open-source DSI implementation supporting all three identifier types, together with a parameterizable and well-documented NQ320K construction pipeline. Experimentally, we achieve results that are competitive with or stronger than previous DSI baselines. Moreover, we conduct extensive experiments under model downscaling, covering retrieval effectiveness, parameter efficiency, training methods and decoding strategies, opening novel directions for future research.