OneSearch: How Kuaishou’s Large‑Model Engine Redefines E‑Commerce Search

Evolution of E‑Commerce Search

Early e‑commerce search relied on simple keyword matching against inverted indexes, returning all items whose titles contained the query terms without understanding user intent. This "dictionary‑lookup" approach produced many irrelevant results and could not personalize recommendations.

With the rise of machine learning around 2010, systems began to incorporate CTR and CVR models, user behavior features, and learning‑to‑rank (LTR) stages. Although these models improved relevance, they required extensive hand‑crafted features (price, brand, color, etc.) and operated in a fragmented three‑stage pipeline (recall → coarse ranking → fine ranking). Information was repeatedly recomputed, leading to high latency, low model‑utilization, and poor handling of cold‑start items.

Why a Generative, End‑to‑End Approach?

As large‑model capabilities matured, researchers questioned the necessity of splitting search into separate stages. If a single model could jointly understand the query, the product, and the user, it could generate the most appropriate results directly.

OneSearch Design Overview

OneSearch replaces the traditional pipeline with a single end‑to‑end model that simultaneously performs user intent understanding, product comprehension, and result generation.

1. Hierarchical Quantization Encoding (KHQE)

Traditional search treats product titles as bags of words, which become noisy when merchants overload titles with keywords (e.g., "French dress summer slim fit commuting evening blue"). OneSearch introduces KHQE, a five‑layer semantic encoding that first clusters products into coarse categories (e.g., "apparel → women → dresses") and then adds two finer‑grained layers encoding attributes such as color, style, and material. Each product is thus represented by a compact vector containing hierarchical semantic information, reducing noise from keyword stuffing.

2. Integrated User Profiling

Instead of feeding sparse, hand‑crafted user statistics to later stages, OneSearch injects the entire user behavior sequence (recent searches, clicks, adds‑to‑cart, purchases) into the model from the first inference step. Short‑term actions receive higher weight, while long‑term preferences are distilled into a few dense vectors via a sliding‑window augmentation, allowing the model to capture both immediate interests and stable tastes without increasing latency.

3. Preference‑Aware Ranking System (PARS)

Traditional ranking treats relevance and personalization as separate signals, often applying preference features only in the final stage. OneSearch’s PARS merges relevance and preference into a unified training objective. The model first learns semantic matching between queries and products, then learns co‑occurrence patterns, and finally incorporates user profiles. Real‑world user actions (clicks, adds‑to‑cart, orders, dwell time) serve as hierarchical reinforcement‑learning rewards, continuously aligning the model’s ranking with current user preferences.

Performance Impact

Production metrics reported by the Kuaishou team show a jump in model‑utilization (MFU) from 3.26 % to 27.32 %, cutting inference cost to one‑quarter of the previous system. Online A/B tests recorded a 1.67 % lift in single‑item CTR, 3.14 % increase in page‑level CTR, 1.78 % rise in CVR, a 2.4 % growth in buyer count, and a 3.2 % boost in order volume. Improvements were even larger for long‑tail queries and cold‑start items.

Conclusion

OneSearch demonstrates that replacing the decades‑old multi‑stage search pipeline with a unified large‑model architecture can simultaneously improve efficiency and effectiveness, marking a significant shift in the direction of e‑commerce search technology.