Scaling Hybrid Retrieval from k=100 to 2,000 Lifts Pool Recall to 0.871
CShorten30 · x · 2026-11-16
Part 3 of the Scaling Reranked Retrieval thread: hybrid retrieval (BM25 + Snowflake Arctic 2.0 embeddings) is scaled from k=100 to k=2,000, with average pool recall climbing steeply from 0.476 to 0.871.
This recall is the ceiling every downstream stage inherits — a relevant document not in the pool cannot be recovered later.
More from Research
- Jon Barron's CVPR23 Talk Revisited: A 1950s Speed Forecast That Overestimated the Future — jon_barron · 2026-11-16
- 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