Answers are grounded in this paper and cite their sources.
Retrieval-augmented generation grounds a language model's output in an external corpus, but flat similarity search over chunks struggles on multi-hop questions that require combining a fact in a table with one stated only in a linked passage. GraphRAG addresses this by extracting an explicit entity-relationship graph before answering, but it is not obvious that this helps every underlying model equally, since building and reading the graph depend on capabilities that vary across models. This article evaluates GraphRAG against a strong hybrid dense/sparse retrieval baseline across four open-weight language models on 200 HybridQA questions, and extends this into a full 4 × 4 index-model/generation-model ablation that independently varies which model builds the graph index from which model queries it. Mismatching the index- and generation-model collapses raw exact match by 27 percentage points, but this collapse is predominantly a citation-style formatting artefact: a mismatched generation model begins reproducing forbidden internal provenance identifiers in its output, exposing index structure to end users, rather than losing factual accuracy outright. Swapping the generation model for cost or deprecation reasons therefore silently reconfigures the system's disclosure surface without rebuilding the persistent, expensive graph index. These findings are actionable for privacy-constrained deployments reliant on self-hostable, open-weight models: GraphRAG's benefit should not be assumed to transfer across models, cost is not a safe proxy for index quality, and citation-style formatting drift should be monitored as a compliance signal. Code is available at https://github.com/HassanMuhammadSannaullah/thesis-rag-graphrag.
No takes yet. Share an insight, caveat, or question.
Muhammad et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: