Problem
Search ranking merges many signals into one order: match to the query, business value such as popularity or availability, and user signals such as preferences and geography. The usual reflex is to let the business set numeric weights (boosts). Nobody can tell whether those weights are the best ones, the best values differ by query type, and every setting needs its own test period. Teams also tune weights on a handful of queries without checking the global effect.
Example from the article: sorting a store's results by increasing price returns items that are cheap but far from the query, because the engine scores every potential hit, even distant ones. Users then have to add filters.
Solution in brief
Replace numeric weights with structure the business can state with confidence: an ordered list of attributes instead of boosts, and relevance filtering or buckets before any business sort. Let a learning-to-rank model learn the merge of text, business and user signals from behavior.
Questions this case answers
- How to merge heterogeneous ranking signals into one order when a single hand-set weight is applied equally to every candidate, including the irrelevant ones
- How to express business knowledge about field importance to a ranker without fixing numeric boosts
- How to combine textual, business and user signals in search ranking without hand-tuned weights
- How to stop a sort by price or popularity from filling the top with irrelevant items
Tools and methods
- Ordered attribute hint — attributes ranked by importance, from which an algorithm learns weights.
- Relevance buckets — textual relevance tiers; the business score sorts only inside a tier.
- Learning to rank (LTR) — a model, often boosted trees, that merges signals per context.
- BM25F — keyword textual score, a typical input.
- Cross-encoder re-ranker — merges keyword and vector hits.
Google or a one-time email code. No password.