From hackathon to MVP: Building a smarter beauty search

Syscily Brown10 Feb 2026

Problem

Shoppers of skincare products start from a concern ("dry skin", "dark spots"), not from a product type ("serum", "face wash"). A category-first search makes them guess the category. It also does not say which ingredients help or whether the recommendation can be trusted. A wrong match means the buyer pays for a product that does not fit the concern.

Example from the article: a shopper types "dry skin" and needs products with the ingredients that address it, plus a plain-language reason why they were shown.

Solution in brief

Make concerns and ingredients first-class searchable fields, add synonyms for beginner and technical terms, and encode a product philosophy into ranking: more active ingredients first, fillers pushed down, ratings as a tiebreaker. An LLM only explains the top result (why it matched, star ingredients, cautions) and never retrieves or ranks products.

Questions this case answers
  • How to rank catalog items for a need-based query when the item taxonomy does not match the words users search with
  • How to add generated explanations to search results while keeping retrieval and ranking deterministic
  • How to make product search answer a shopper's concern (dry skin, dark spots) when the catalog is organized by product type and ingredients
  • How to let beginners and experts find the same products when they use different vocabulary

Tools and methods

  • Algolia — indexing, retrieval, searchable attributes, synonyms, ranking, tuning in the dashboard
  • Gemini — explanation layer from concern, ingredient list and a requested answer structure
  • React — custom hit grid and a search connector with pagination
  • Prompt engineering — fixed answer structure: benefits, star ingredients, cautions, tips
Continue reading — it's free

Google or a one-time email code. No password.