CAGE: Enhancing RAG via Coherence-Aware Graph Encoding
September 7, 2026
CAGE improves RAG reranking by modeling coherence between retrieved passages as a directed heterogeneous entity graph. Using a Relational Graph Convolutional Network, it optimizes for informational bonding and factual consistency, matching or exceeding monoT5 on multi-hop benchmarks.
HOW THIS AFFECTS YOU
●
builderYou can reduce noise in RAG pipelines by prioritizing structurally coherent context sets over isolated relevant chunks.
●
researcherYou can evaluate RAG effectiveness through multi-dimensional between-chunk coherence metrics.