When data is read far more often than written, you can afford to organize it once and reuse that work on every query.

Sub-millisecond hybrid search (dense + sparse), a from-scratch HNSW index, BM25 lexical ranking, reciprocal-rank fusion, feature re-ranking and a complete retrieval-augmented-generation pipeline — all ...

Each row points at its in-game proof (headless, BMG G21) and its benchmark (tools/bench_create.py, benchmarks/); the phase table below has the detail, and "Blocked / open" what isn't done. Numbers ...