What this is
Dewu, the Chinese sneaker and streetwear authentication and trading platform, recently published a long technical article detailing a complete rebuild of the recall stage in transaction search. Recall is search's first gate: pick a few hundred candidates out of hundreds of millions of products, then hand them off to a ranking model for fine ordering. The traditional approach is 'discriminative'—either match keywords, or convert both products and search queries into vectors (numeric strings) and compute similarity. Dewu identifies two unavoidable pitfalls: a user searching for the Nike Air Force 1 sneaker (colloquially '空军一号' in Chinese, literally 'Air Force One') while the product title reads 'AF1'—a complete literal mismatch; and new cold-start products with no user behavior data, where vector models also fail to retrieve accurately. So Dewu switched tack: use large language models (AI that can understand and generate text) to 'generate' tags for products—extract visual features like 'sheer' or 'vintage' from images, mine colloquial expressions from user reviews, use tags from related products to enrich sparse items, run OCR on hidden text in detail-page images like '100W fast charging', and finally anchor high-frequency search terms back into the index.
Industry view
We note this is the first time a top-tier Chinese e-commerce team has systematically disclosed a complete engineering chain for 'generative recall'—worth close study by peers. But several points warrant cooler heads: First, the models cited—Qwen3-32B and Qwen2.5-VL-7B (a vision-language model that can 'see and speak')—carry nontrivial inference costs. Each product tag requires several rounds of LLM inference, and at scale costs could run tens of times higher than traditional vector retrieval. Second, the three-layer defense of 'model filter + rule filter + manual review' sounds solid, but leakage rates and long-term maintenance costs are not disclosed. Third, the entire pipeline is highly dependent on stable output from the Qwen model family; version updates may require recalibration. We believe the more prudent direction is likely a 'discriminative + generative' hybrid architecture—not a wholesale replacement.
Impact on regular people
For e-commerce IT teams: this is a construction manual for 'rebuilding search with LLMs,' but before copying the homework, you must calculate the inference cost ledger—you can't just look at effectiveness.
For working professionals: once search quality improves, operations staff, buyers, and others whose roles depend on 'sourcing goods' will find that long-tail products get surfaced more often.
For the consumer market: over the next 1-2 years, search experiences on mainstream e-commerce platforms will see visible improvements—especially for new releases, niche items, and products you 'can't quite name but know what they look like'.