Problem
Online stores need each product card to carry attributes such as product type, volume in ml and suitable skin type. Gaps hurt filters, search and sales. Filling them by hand from manufacturer data is expensive. Matching to similar products and copying values fails when many products are unique. Yandex Crowd Solutions faced this for several cosmetics and perfumery stores, which also required up-to-date values from manufacturer or near-official sites.
Example from the article: a "volume" attribute was normally in milliliters, but for sheet masks some sites wrote "1 (one piece)". The model read "1" as milliliters and reported that a sheet mask contains 1 ml.
Solution in brief
Route each attribute by type to the cheapest source that can answer it. A search-enabled LLM fetches facts, a second LLM extracts the value in the required format and maps it to the allowed list. A vision model (VLM) reads visual attributes from the product photo. Subjective attributes are derived by an LLM from a product description. Critical attributes still go to people.
Questions this case answers
- How to fill missing attributes of catalog entries from web search results and images, routing each attribute to the source that can answer it and mapping values to an allowed list
- How to decide which catalog attributes an LLM can fill without review and which must go to human review
- How to fill missing product attributes in an e-commerce catalog with LLMs
- How to extract attribute values from web search results in a fixed JSON format and map them to an allowed value list
Tools and methods
- Neuro (YandexGPT-based LLM with web search) — finds attribute values in web sources
- YandexGPT (second LLM) — extracts the value from the raw answer and maps it to a closed list
- VLM — determines bottle color, cap color and bottle and cap features from the image
- Human reviewers (Yandex.Tasks crowd) — recheck critical and hard cases
Google or a one-time email code. No password.
