Our summary
Finding a relevant page is only part of search. The first results may repeat the same product, or a document may be split halfway through an answer. Qdrant's experiments ask where Jev can help: ordering results, adding variety, choosing categories and deciding where a new topic starts.
One method scores the search results, then combines those scores with a rule that rewards variety. Another boosts a likely product category and filters only above a higher cutoff. For documents, Jev judges gaps between sentences; ordinary code uses those judgments to make chunks within size limits.
The authors report gains on their chosen datasets, but also show an apple search losing electronics after a confident food-category guess. Once filtered out, a useful result cannot be recovered by reordering. Proposed fixes for short queries remain untested, and hosted checks send queries or document text to OpenRouter.
Key takeaways
- A relevance ranking can still show near-duplicates. Test the separate rule that gives different results room to appear.
- Try ambiguous queries before filtering by category. A boost changes order; a filter removes candidates that later ranking cannot recover.
- Split documents before a search request arrives when possible, so that work does not add a model call to each search.