Open-Source Reproduction and Explainability Analysis of Corrective Retrieval Augmented Generation

A new research paper presents a fully reproducible, open-source implementation of Corrective Retrieval Augmented Generation (CRAG), a system designed to make AI more reliable by evaluating the quality of documents it retrieves before generating an answer. The original CRAG system relied on proprietary components like the Google Search API and closed model weights from LLaMA-2, making it difficult for other researchers to study or build upon. This work, by Surya Vardhan Yalavarthi, successfully replaces those components with open alternatives: the Wikipedia API for search and Microsoft's compact Phi-3-mini-4k-instruct model as the generator.

Crucially, the open-source pipeline achieves performance comparable to the original system on standard benchmarks like PopQA and ARC-Challenge, proving the core CRAG architecture can be effectively replicated without closed tools. Beyond replication, the paper provides the first explainability analysis of CRAG's T5-based retrieval evaluator using SHAP (SHapley Additive exPlanations). This analysis reveals that the evaluator judges document relevance primarily by checking for named entity alignment—matching proper nouns like people and places—rather than deeper semantic understanding.

The research also identifies key failure modes, notably that the system struggles with domain transfer, particularly on science questions where entity matching is insufficient. All code and results are publicly available, providing a crucial resource for the community to audit, improve, and build more transparent and robust RAG systems. This work demystifies an advanced AI technique and provides a blueprint for creating reproducible, explainable AI pipelines.