Researchers have proposed a new approach to improve the efficiency of graph-based retrieval-augmented generation (RAG) in long-video understanding. They argue that existing systems unnecessarily couple the granularity of indexing with evidence reasoning. To address this issue, they introduced Density-Aware Graph Construction (DAGC), a training-free method that decouples coarse retrieval indexes from fine-grained evidence spaces. DAGC constructs a compact graph index by merging visually redundant chunks while preserving mappings to original temporal units. Experiments show that DAGC achieves significant speedup and performance retention on various long-video benchmarks. --- Why it matters: This matters because it could enable faster and more efficient processing of long videos, which is crucial for applications such as video understanding and retrieval in areas like surveillance, education, and entertainment. Source: https://arxiv.org/abs/2608.23011