Semantic search

How this question breaks down

1. 1QueriesCore19 minTurn the few words someone typed into something the retrievers can use. Fix the spelling, guess the category, pull out sizes and brands, and rewrite queries that would otherwise return nothing.
2. 2Dense retrievalCore15 minEmbed queries and listings with a text bi-encoder fine-tuned on Boltwise's own searches, so "stop door slamming" finds a soft-close damper that shares none of its words.
3. 3HybridCore19 minRun keyword and dense retrieval side by side and merge their lists, with a weight that leans on keywords for part numbers and on meaning for descriptions.
4. 4RankingCore18 minOrder the ~300 merged candidates with a gradient-boosted ranker over relevance, quality and delivery features, then let a cross-encoder read the query and the top 40 listings together.

Work through the parts in order and you’ve covered the full design. Each part is one topic.