Scaling Reranked Retrieval: Hybrid Search + LLM Reranking Lifts nDCG@10 2.6x
CShorten30 · x · 2026-11-16
Christopher Shorten shares a research report, Scaling Reranked Retrieval, framed around three levers for spending the next unit of search compute: depth (deeper retrieval), stages (more reranking stages), and width (more models per stage).
Depth and stages compose into a quality ladder with nDCG@10 climbing at each rung:
- 0.181: hybrid retrieval alone (BM25 + Snowflake Arctic 2.0 embeddings)
- 0.358: + zerank-2 reranking the top-100
- 0.424: + a deeper pool, reranking the top-2K
- 0.470: + GPT-5.6 Luna listwise pass over the top 20
More from Research
- Google opens SatWeave and PARDIG, the foundational datasets behind WeatherNext 3 — ymatias · 2026-11-16
- Anthropic: agent memory works best by forgetting selectively—notes beat 400+ archives 4x in Pokémon Red — The Decoder · 2026-11-16
- Plasma proteomics outpredicts HbA1c for 10-year cardiometabolic disease risk in Nature Metabolism study — EricTopol · 2026-11-16
- Hugging Face ships definitive fine-tuning guide and scripts for 3D model TRELLIS.2 — linoy_tsaban · 2026-11-16
- miHoYo's Gamer agent beats 95% of Honkai: Star Rail challenges, plays 30+ games — _akhaliq · 2026-11-16
- Lightwheel launches RoboFinals RL-Env: RL training envs from real industrial tasks with calibrated physics — chris_j_paxton · 2026-11-16