Semantic product index
MiniLM embeddings represent product names, descriptions, types, colours, garment groups, and departments. Field-weighted ranking returns products related by meaning, not only by exact word matches.
H&M product discovery
Search for fashion in the way people actually shop.
A product-discovery interface that blends semantic search, customer taste profiles, and LLM-assisted shopping chat.
Try the live search & chat experience ↗Built from the H&M Kaggle dataset ↗The experience
Fashion searches are descriptive and subjective: “bags for overnight travel”, “something for a beach holiday”, or “a red dress for dinner”. Keyword matching alone does not reliably understand those intents.
This project turns H&M’s product catalogue into an interactive discovery experience, with a product grid, images, price filters, suggestions, and an optional customer ID for personalisation.
How it works
Personalised discovery
Purchase histories become recency-weighted taste vectors. Search relevance and customer taste are blended to make results feel individual while preserving the shopper’s stated intent.
LLM shopping chat
A server-side LLM interprets conversational requests and maps them to the semantic search layer. The user sees products, not raw model output; sensitive keys remain on the server.
Connection to the recommender
The same data, two useful experiences.
This interface helps a shopper discover products now. The companion recommendation project predicts what they may purchase next—using an offline MAP@12 evaluation workflow.
Explore the recommendation project →