Alibaba-style CQ-SID framework positions generative retrieval as a recall-stage supplement in e-commerce search
A paper describes CQ-SID, a generative retrieval method for large product catalogs, paired with a reinforcement learning step that aligns recall with ranking goals.
In plain words
Generative retrieval replaces several separate search stages with one model that generates item identifiers directly. The paper's CQ-SID method compresses items into short semantic codes so the model searches faster, while its EG-GRPO training step nudges results toward what the ranking system prefers. The authors report gains in offline tests and online A/B trials on Tmall app search.
What was announced
The authors argue that generative retrieval, which unifies multi-stage retrieval into a single end-to-end model, has been hard to adopt in industrial e-commerce because catalogs are massive and dynamic, latency budgets are strict, and retrieval must serve downstream ranking. Rather than replacing the existing pipeline, they position their method as a recall-stage supplement. CQ-SID encodes items into hierarchical semantic cluster identifiers using category-aware and query-item contrastive learning plus Residual Quantized VAEs, which they say cuts beam search complexity.
On offline experiments built from Tmall app search logs, the paper reports up to 26.76% and 11.11% relative gains in semantic and personalized click hitrate over RQ-VAE baselines, with beam search size halved. EG-GRPO, a reinforcement learning method that injects ground-truth samples to stabilize training under sparse rewards, is reported to improve multi-objective performance. Online A/B tests showed GMV up 1.15% and UCTCVR up 0.40%. The authors state the generative recall channel now handles over 50.25% of exposures, 58.96% of clicks, and 72.63% of purchases in their production system. These figures come from the paper alone and have not been independently verified.
What it does not show
- Evidence is limited to the paper's abstract; full methodology, baselines, and test conditions are not available for verification.
- Reported metrics and production shares are the authors' own claims; no independent confirmation or comparison with other published systems is provided.
For security and IT teams
- Status
- The paper states the generative recall channel is already in production, contributing the reported shares of exposures, clicks, and purchases.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.