SproutRAG: Attention-Guided Tree Search with Progressive Embeddings for Long-Document RAG

SproutRAG is a novel attention-guided hierarchical retrieval-augmented generation (RAG) framework designed to enhance the retrieval granularity and contextual coherence of long documents. It organizes sentence-level chunks into progressively larger semantic units using learned inter-sentence attention, allowing for multi-granularity retrieval without incurring additional LLM calls or lossy summarization. Experimental results show a 6.1% improvement in information efficiency across various benchmarks, making it a significant advancement for practitioners seeking to optimize retrieval processes in complex document scenarios.