Retrieval-Augmented Generation (RAG) mitigates factual hallucinations in Large Language Models but struggles with complex reasoning. Graph-based RAG (GraphRAG) addresses this gap yet can exhibit a failure pattern: semantic compression during triple extraction may reduce the preservation of fine-grained facts such as precise quantities, temporal constraints, and entity attributes. We formalize this limitation as the Fine-Grained Fact Retrieval (FGFR) cascade and introduce stage-wise diagnostic metrics, including R_rep, R_sel, and R_end, to quantify information loss. To stress-test FGFR, we construct a diagnostic evaluation suite from classic novels that provide controlled long-context scenarios containing explicit entity attributes and temporal constraints. F-GraphRAG, our framework, features a granularity-aware extractor that preserves atomic details and an information-preserving filter that counters global importance bias. On this evaluation suite, F-GraphRAG outperforms GraphRAG by 12.6% in accuracy while using 33.8% fewer relationships, and achieves gains of 5.3–11.2% across nine public benchmarks. These results suggest that improving evidence completeness can provide a more effective alternative to increasing graph connectivity.