Rethinking RAG in Long Videos: What to Retrieve and How to Use It?
Researchers have introduced V-RAGBench, a new benchmark designed to evaluate retrieval-augmented generation (RAG) systems specifically for long videos. This benchmark addresses limitations in existing methods by creating query-relevant evidence chunks and enabling decoupled evaluation of retrieval and generation. Additionally, a new method called CARVE is proposed, which utilizes parallel retrievers and chunk-adaptive reranking to select the optimal configuration for each video chunk, improving performance over existing VideoRAG baselines. AI
IMPACT This research could lead to more accurate and nuanced AI understanding of video content, improving applications that rely on video analysis.