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

Short summary

SproutRAG presents an attention-guided hierarchical framework for RAG systems that organizes documents into semantically coherent units without expensive LLM calls. Using learned inter-sentence attention and hierarchical beam search, it achieves 6.1% better information efficiency across scientific, legal, and open-domain benchmarks. Code is available on GitHub.

• •Hierarchical RAG framework using attention mechanisms to build multi-granularity retrieval trees
• •Eliminates costly LLM calls during indexing while preserving full context without summarization
• •6.1% average improvement over baselines across four benchmarks with open-source code

Generated with AI, which can make mistakes.