Retrieval · Ablation · Hybrid

Twenty-nine thousand toys, one query away.

A vision-language model tags every product. Those tags are clustered into 4,609 canonical labels. A 64-dimensional autoencoder re-ranks what the text filter catches. The result — a 4.45× lift over the naïve baseline — is shown live for eight held-out queries below.

Precision @ 10: 0.154; hybrid mean · 132 queries
Lift over baseline: 4.45×; vs. autoencoder only
Catalogue: 29,637; SKUs · 99.98% scraped
Build cost: < HK$500; one-off API spend

§ I

Pick a query

Or use the list

§ II

Dispatch

§ III

Top-ten findings

Hybrid retrieval · text-tag filter → 64-d embedding re-rank. Ground-truth hits are marked.