Does Cross-Encoder Reranking Close the Multi-Hop Performance Gap Between Knowledge Graph Retrieval and Vector RAG in Renewable Energy Domain QA?

Graph RAG is widely assumed to outperform vector retrieval on multi-hop questions, yet most graph pipelines retrieve an unordered candidate set and truncate it, discarding the bridging chunk that a multi-hop question depends on. This work isolates that failure mode. Three pipelines are built over a single knowledge graph such that consecutive pipelines differ in exactly one variable: the reranked variant differs from the unranked graph baseline only in scoring the full candidate pool before truncation rather than after. This separates ordering-before-truncation from cross-encoding itself, an ablation absent from published work, which applies reranking only to candidate lists that already arrive ranked. Single-hop and multi-hop results are reported separately with bootstrap confidence intervals and paired Wilcoxon signed-rank tests. Reranking closes the multi-hop gap to statistical indistinguishability but leaves a significant single-hop gap, and the gain costs 1.8 s of local compute with no additional model calls