Source-Aware Rerouting Improves RAG Retrieval Reliability

A recent study conducted by the Milwaukee School of Engineering presents a straightforward enhancement to Retrieval-Augmented Generation (RAG) frameworks by integrating source credibility into the ranking of documents. Traditional RAG systems evaluate retrieved documents based solely on semantic similarity, neglecting their origins. The new approach assigns a reliability prior, lambda(s), to each document according to its source type and adjusts similarity scores using the formula score(q, d) = sim(q, d) * lambda(s). Evaluated on a health-domain corpus of 120 documents, this source-aware reranking increased Precision@5 from 0.48 to 0.72 and minimized adversarial document retrieval in scenarios where low-credibility sources can be identified through metadata. The experiments utilized Rosie, the Milwaukee School of Engineering's high-performance computing cluster.