Search Engineer | Retrieval, Ranking, and AI-powered Discovery
Search Platform, at Coursera, 12+ years of industry experienced · Search, Relevance & Distributed Systems · Java · OpenSearch/Elasticsearch · AI-driven discovery
Markdown supported · ⌘↵ to post
Kartik N V J K
Your point that nearest-neighbour always returns something, so a no-answer query still surfaces a confident wrong result, is what most teams underestimate. I have found fusing BM25 with dense retrieval only half fixes it; the reranker score distribution is what I threshold on to abstain when nothing clears the bar. How do you decide when to return nothing rather than the top fused hit