Improving search ranking with chess Elo scores
Published:Jul 16, 2025 14:17
•1 min read
•Hacker News
Analysis
The article introduces new search rerankers (zerank-1 and zerank-1-small) developed by ZeroEntropy, a company building search infrastructure for RAG and AI Agents. The models are trained using a novel Elo score inspired pipeline, detailed in an attached blog. The approach involves collecting soft preferences between documents using LLMs, fitting an Elo-style rating system, and normalizing relevance scores. The article invites community feedback and provides access to the models via API and Hugging Face.
Key Takeaways
- •ZeroEntropy released new search rerankers: zerank-1 and zerank-1-small.
- •Models are trained using an Elo-inspired pipeline.
- •One model is open-source.
- •Models are accessible via API and Hugging Face.
Reference
“The core innovation is the use of an Elo-style rating system for ranking documents, inspired by chess.”