NeocorRAG: Less Irrelevant Information, More Explicit Evidence, and More Effective Recall via Evidence Chains

Researchers introduce NeocorRAG, a new framework that optimizes retrieval quality in Retrieval-Augmented Generation (RAG) systems by using Evidence Chains, achieving state-of-the-art performance while reducing token consumption by 80% compared to comparable methods. The framework addresses a critical gap where improvements in retrieval metrics don't consistently translate to better reasoning accuracy.

NeocorRAG tackles a fundamental inefficiency in modern AI systems: the disconnect between retrieval performance and actual reasoning quality. Traditional RAG approaches struggle with a false dichotomy—methods optimizing for high recall often sacrifice quality, while quality-focused approaches suffer from lower recall rates. The researchers quantify this problem through the Recall Conversion Rate metric, revealing that mainstream methods experience near-linear performance decay as retrieval improves, indicating wasted computational resources.

This advancement builds on years of RAG development, where systems augment language models with external knowledge to improve factuality and reasoning. The introduction of Evidence Chains represents a meaningful architectural innovation, allowing the framework to systematically mine and organize supporting information. The activated search algorithm and constrained decoding mechanisms ensure that retrieved information directly supports downstream reasoning tasks.

The efficiency gains matter significantly for practical deployment. Consuming less than 20% of tokens while achieving superior performance on multiple benchmarks (HotpotQA, 2WikiMultiHopQA, MuSiQue, NQ) addresses real concerns about computational costs in AI inference. The training-free nature of the approach enhances accessibility for developers integrating RAG into production systems.

For the AI industry, this research demonstrates that careful architectural design can overcome efficiency-quality trade-offs that seemed intractable. Organizations building AI applications stand to benefit from reduced computational overhead without sacrificing accuracy. The open-source release accelerates adoption across the community. Future work likely focuses on scaling these principles to even larger models and exploring how Evidence Chains apply to other reasoning-intensive tasks beyond question-answering.

• →NeocorRAG achieves state-of-the-art RAG performance while using 80% fewer tokens than comparable methods.
• →The framework introduces Evidence Chains to systematically optimize retrieval quality while maintaining high recall.
• →A new Recall Conversion Rate metric reveals that improved retrieval metrics don't automatically improve reasoning accuracy in existing methods.
• →The training-free approach enables immediate adoption without requiring model fine-tuning or retraining.
• →Results across multiple benchmarks demonstrate consistent improvements on both 3B and 70B parameter language models.