Madhava Hybrid v6: Clustered Index + Bounded Search (200K Structured Benchmark)

# Winnex Madhava Hybrid v6: Clustered Index + Bounded Search ## Architecture Level 1: IVF clustering (64 cells). Level 2: Madhava 32D->64D per cell. Eliminates O(N) linear scan while preserving mathematical bounds. ## Results (200K docs, 50 clusters, 128D, 200 queries) | Method | NDCG@10 | Recall@10 | Latency | Build | Bound | |---|---|---|---|---|---| | FlatIP (exact) | 0.893 | 0.874 | 13.5ms | - | - | | HNSW(ef=128) | 0.910 | 0.893 | 0.74ms | 111s | None | | IVF(nprobe=5) | 0.985 | 0.981 | 0.23ms | <1s | None | | Madhava (1-cell) | 0.879 | 0.857 | 34.9ms | 0.27s | Zero | | **MadHybrid(np=3)** | **0.938** | **0.928** | **1.91ms** | **4.3s** | **Zero** | | MadHybrid(np=5) | 0.919 | 0.905 | 2.76ms | 4.6s | Zero | ## Links Kaggle: https://www.kaggle.com/code/kleniopadilha/madhava-hybrid-clustered-index-v6 GitHub: https://github.com/klenioaraujo/winnex-madhava ## License: BSL 1.1 | pay@winnex.ai