Problem
Yandex Market users often re-sort search results by price. In 90% of those cases they choose ascending price, and sorting accounts for about 10% of orders and add-to-carts. Under sorting the only control left is the filter, because the price order cannot be broken afterwards. A constant score threshold separates relevant from irrelevant items badly: a high threshold cuts relevant items and lowers orders, a low one lets cheap accessories fill the top. Offline nDCG under sorting was about half of that of the default results.
Example from the article: for the query "Garmin watch", the default results start with watches, while after sorting by ascending price the top shows cords and straps.
Solution in brief
Add a second, dynamic filtering stage after the constant threshold. Calibrate filter scores to probabilities, discount them by price position, and pick the subset of price-sorted documents that maximizes DCG with dynamic programming. Also train the sorting model on a large pool labeled by an LLM instead of crowd annotators.
Questions this case answers
- How to keep a ranked result list relevant when the user imposes a fixed attribute order that cannot be changed afterwards, so that filtering is the only lever
- How to choose the subset of candidates to show that maximizes expected discounted gain when raw relevance scores are uncalibrated and low-attribute items are systematically under-scored
- How to keep search results relevant when the user sorts them by price
- How to filter irrelevant documents without losing recall of cheap relevant items
Tools and methods
- Dynamic filtering — picks the DCG-maximizing subset of the price-sorted candidates.
- Dynamic programming — finds that subset without enumerating all options.
- Sigmoid calibration — turns the raw filter score into a probability-like relevance value.
- Price-position discount with — protects cheap relevant items.
- YandexGPT labeling — replaces Toloka annotators for the training pool.
Google or a one-time email code. No password.
