Ψ-RAG: Hierarchical Abstract Tree for Cross-Document Retrieval
Researchers propose Ψ-RAG, a tree-based retrieval-augmented generation framework designed to handle cross-document multi-hop queries. Existing Tree-RAG methods, originally built for single-document retrieval, suffer from poor distribution adaptability due to k-means clustering noise, structural isolation lacking cross-document connections, and coarse abstraction that obscures fine details. Ψ-RAG addresses these with a hierarchical abstract tree index built via iterative merging and collapse, adapting to data distributions without prior assumptions, and a multi-granular retrieval agent that intelligently interacts with the index. The framework aims to scale RAG to complex queries spanning multiple documents.
Key facts
- Ψ-RAG is a tree-RAG framework for cross-document retrieval
- Existing Tree-RAG methods face challenges in scaling to multi-hop questions
- Problems include poor distribution adaptability, structural isolation, and coarse abstraction
- k-means clustering introduces noise due to rigid distribution assumptions
- Tree indexes lack explicit cross-document connections
- Ψ-RAG uses a hierarchical abstract tree index built through iterative merging and collapse
- The index adapts to data distributions without a priori assumptions
- A multi-granular retrieval agent interacts with the index
Entities
Institutions
- arXiv