How OneSug Revolutionizes E‑commerce Query Suggestion with End‑to‑End Generative Modeling

Research Background

In e‑commerce search a single term (e.g., “apple”) can express multiple intents. Accurate query suggestion is essential for user experience and conversion. Traditional multi‑stage cascade architectures (recall → coarse‑ranking → fine‑ranking) suffer from mismatched objectives, difficulty recalling long‑tail queries, and limited overall performance.

OneSug: End‑to‑End Generative Framework

OneSug unifies recall, coarse‑ranking and fine‑ranking into a single generative encoder‑decoder model, improving recommendation quality and system efficiency.

Key Components

Prefix‑Query Representation Enhancement (Prefix2Query) : A base BGE model is fine‑tuned on real prefix‑to‑query and query‑to‑query pairs using contrastive learning to align the semantic space with e‑commerce data. An RQ‑VAE maps any text to hierarchical semantic IDs, enabling fast top‑K matching of queries.

Unified Encoder‑Decoder Architecture : The model generates the most likely clicked query autoregressively. Input consists of (1) current user prefix, (2) enhanced query sequence from Prefix2Query, (3) user historical behavior sequence, and (4) user profile features.

User Preference Alignment (Reward‑Weighted Ranking, RWR) : User search behavior is divided into six levels with base reward weights; a CTR‑based adjustment factor is applied. Nine sample‑pair types (e.g., <Order, Show>, <Click, Rand>) are constructed, and a weight inversely proportional to the reward gap is assigned—large gaps receive smaller weights (easy samples), small gaps receive larger weights (hard samples).

Mixed Ranking Objective

OneSug combines pairwise and listwise losses. The listwise loss (inspired by Plackett‑Luce) pushes the reward of a positive sample above all negatives, while a point‑wise SFT loss preserves generation quality and prevents reward hacking.

Experimental Evaluation

Offline experiments on a large‑scale Kuaishou e‑commerce dataset show that OneSug outperforms traditional multi‑stage pipelines and generative baselines on HR@16 and MRR@16. Online A/B tests report significant lifts in CTR, order volume and GMV, and a 43.2 % reduction in average latency after replacing the cascade with the unified model.

Paper:

https://arxiv.org/abs/2506.06913

Conclusion

OneSug is the first fully deployed end‑to‑end generative query suggestion system in e‑commerce, delivering gains in semantic understanding and personalization. Future work includes exploring large language models for ranking reinforcement learning, real‑time model updates, and extending the approach to advertising and other scenarios.