Natsuki

Hybrid BM25 + dense retrieval search engine built from scratch (no IR library dependency), with a LambdaMART learned reranker and LLM query expansion on top. Validated end-to-end on BEIR/SciFact against real relevance judgments — each stage (bm25 < dense < hybrid < rerank) shows a clean, honestly-measured improvement, including an A/B test where query expansion was found to hurt more than it helped.