Kuaishou's OneRetrieval unifies e-commerce search with editable generative AI

Industrial e-commerce search typically relies on multiple retrieval branches (e.g., inverted index, vector search, and learning-to-rank) fused by hand-tuned heuristics. This multi-branch setup is hard to optimize jointly and inefficient. Generative retrieval (GR) promises a single-model alternative, but existing GR methods cannot match the real-time editability of the inverted-index branch — the only branch that allows operators to inject new terms within hours without any model update.

OneRetrieval solves this with Keyword-Aligned Encoding (KAE), which maps each identifier position to an interpretable attribute word, blending competitive recall with the editability of traditional inverted indexes. The model organizes 18 attribute categories into six codebook groups with non-uniform capacity and reserves slots for binding new words after deployment — no retraining needed. A four-stage fine-tuning pipeline jointly optimizes quality and editability. On 5 million real e-commerce queries, OneRetrieval matches the deep recall of the strongest generative baseline, while its intervention hit rate is an order of magnitude higher than closed-codebook methods. In production at Kuaishou — serving hundreds of millions of page views daily — replacing the inverted-index branch alone significantly lifts order volume, and extending the model to nearly the entire retrieval stage holds conversion while improving CTR.