Idea
An open-weight reranker training method improving retrieval accuracy across domains for enterprises and AI developers.
Research Paper
Core Innovation
This paper introduces zELO, a training method inspired by ELO rating and Thurstone models to optimize ranking tasks. It leverages unsupervised data to train reranker models that outperform proprietary alternatives. The approach enables efficient end-to-end training from unannotated data, achieving strong zero-shot generalization.
Market Size (TAM)
$2–10B TAM for AI-powered search and retrieval; $1–2B SAM from enterprise search in finance, legal, and STEM sectors. Driven by demand for improved search accuracy and open-source AI models.
Potential Customers & Pain Points
- Enterprises needing high-accuracy document retrieval
- AI developers lacking open-source reranker models
- Legal and finance firms requiring domain-specific search
- Code repositories seeking better code search relevance
- STEM researchers needing robust out-of-domain retrieval
Business Model
Offer open-weight reranker models with enterprise support subscriptions and custom training services.
Competitive Landscape
- OpenAI
- Google AI
- Microsoft Azure Cognitive Search
Implementation Challenges
- Competition from large closed-source models
- Need for extensive computational resources
- Adoption inertia in enterprise environments
Validation Strategy
- Benchmark against proprietary rerankers on multiple domain datasets
- Deploy pilot projects with finance and legal firms
- Collect user feedback on zero-shot and domain-specific performance
Research Paper Overview
zELO: ELO-inspired Training Method for Rerankers and Embedding Models
Summary
We introduce a novel training methodology named zELO, which optimizes retrieval performance via the analysis that ranking tasks are statically equivalent to a Thurstone model. Based on the zELO method, we use unsupervised data in order train a suite of state-of-the-art open-weight reranker models: zerank-1 and zerank-1-small. These models achieve the highest retrieval scores in multiple domains, including finance, legal, code, and STEM, outperforming closed-source proprietary rerankers on both NDCG@10 and Recall. These models also demonstrate great versatility, maintaining their 0-shot performance on out-of-domain and private customer datasets. The training data included 112,000 queries and 100 documents per query, and was trained end-to-end from unannotated queries and documents in less than 10,000 H100-hours.