InterveneRAG: Diagnosing Where GraphRAG Answers Can Be Recovered

abstract Retrieval-augmented generation (RAG) helps large language models answer knowledge-intensive questions by supplying external evidence. GraphRAG organizes this evidence into graphs for questions requiring facts from multiple sources. Incorrect answers may arise from missing graph information, retrieval failures, or ineffective evidence use, but component metrics do not directly show whether a stage repair would improve the answer. We introduce InterveneRAG, a diagnostic framework measuring recoverability: the paired change in answer quality under a specified stage intervention. Annotation-guided construction and retrieval replacements and evidence-ordering probes quantify remaining repair potential with downstream operators fixed. Individual and joint interventions, together with graph-size and context-budget controls, examine how gains depend on stage interactions and evidence selection and presentation. Across two multi-hop question-answering datasets and three generator families, annotation-guided selection within existing automatic graphs improves answer F1 by $0.206$–$0.339$, despite those graphs containing about 95% of annotated supporting facts. Extensions with two generators yield F1 gains of $0.293$–$0.305$ on an equal-hop MuSiQue cohort and $0.086$–$0.117$ through selection restricted to existing passages in bounded shared HippoRAG indices. Construction–retrieval interactions vary across comparison protocols; on 2WikiMultiHopQA, a new factorial protocol combining common budget ceilings with a revised retrieval intervention yields interaction estimates whose confidence intervals cross zero. These findings reveal substantial selection headroom within existing graphs and indices while showing that stage interactions require protocol-specific interpretation. The measured gains characterize annotation-guided repair potential rather than guaranteed deployable improvements. Code: https://anonymous.4open.science/r/InterveneRAG-1E85