Agents across catalogue, search, and support — at a scale where a one-percent error rate is thousands of bad results a day.
Marketplaces are a retrieval problem before they are an agent problem. Catalogues are messy, multilingual, full of near-duplicates, and changing hourly. Most AI search projects here fail on data quality, not on the model.
At marketplace volume, model spend is a product constraint. We design retrieval so the expensive model only runs on the small share of queries that need it.