PoP-RAG: Holistic Query Planning for GraphRAG

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Graph-based retrieval-augmented generation (GraphRAG) leverages knowledge graphs to provide context for large language models (LLMs) to generate quality responses. Yet existing GraphRAG methods suffer from two drawbacks: connecting each entity to all passages that mention the entity causes one-to-many entity-passage mapping problem and retrieves redundant passages, and fixed graph traversal patterns fail to locate target information for hard queries. To tackle the two limitations, we propose PoP-RAG, which features a passage-on-edge (PoE) graph that links the passages with graph edges to resolve the one-to-many entity-passage mapping problem. PoP-RAG further introduces a query planning step that decomposes a query into sub-queries and builds a directed acyclic graph (DAG) to model their relationships. This design enables targeted retrieval for each sub-query and ensures transparent response derivation logic, thus enhancing explainability. Experimental results demonstrate that PoP-RAG outperforms existing GraphRAG methods, especially on complex queries.