NeocorRAG: Evidence Chains Boost Recall in RAG Systems
Researchers have identified a critical gap in Retrieval-Augmented Generation (RAG): improved retrieval performance does not consistently translate to better downstream reasoning. To quantify this, they introduce the Recall Conversion Rate (RCR), a novel metric measuring retrieval's contribution to reasoning accuracy. Analysis of mainstream RAG methods shows that as Recall@5 improves, RCR decays near-linearly, indicating a trade-off between recall and retrieval quality. The team proposes comprehensive retrieval quality optimization criteria and introduces NeocorRAG, which uses evidence chains to reduce irrelevant information and enhance explicit evidence, leading to more effective recall. The work is detailed in arXiv paper 2604.27852.
Key facts
- Recall Conversion Rate (RCR) is a new metric to quantify retrieval's contribution to reasoning accuracy.
- As Recall@5 improves, RCR exhibits near-linear decay in mainstream RAG methods.
- NeocorRAG uses evidence chains to reduce irrelevant information and enhance explicit evidence.
- The research identifies a trade-off between recall and retrieval quality in RAG systems.
- Comprehensive retrieval quality optimization criteria are proposed.
- The paper is available on arXiv with ID 2604.27852.
- The work addresses the gap where improved retrieval does not consistently improve reasoning.
- NeocorRAG aims for more effective recall via evidence chains.
Entities
Institutions
- arXiv