HippoRAG

Description

HippoRAG is a research RAG framework from OSU NLP Group that draws on hippocampal indexing theory: documents are decomposed into entity-relation triples that form a knowledge graph, and retrieval runs Personalized PageRank from question entities across that graph. The shape targets multi-hop (associative) questions and sense-making over large corpora where naive dense retrieval fails. HippoRAG 2 (the current implementation) builds on the same PPR core and adds deeper passage integration and more effective online use of an LLM. It is distributed as a Python library and reference implementation accompanying the NeurIPS 2024 HippoRAG paper and the ICML 2025 HippoRAG 2 paper.

Solution

Two-phase pipeline. Offline indexing: documents are passed through an LLM to extract OpenIE-style entity-relation triples that are merged into a persistent knowledge graph with vector embeddings on nodes. Online retrieval: at query time the system extracts entities from the question, seeds Personalized PageRank from those entity nodes over the KG, and returns the top-scoring passages associated with the highest-ranked nodes for an LLM reader (the rag_qa step in the Python API).

Primary use cases

• multi-hop (associative) question answering over a document corpus
• knowledge-graph-augmented retrieval pipelines
• sense-making over long / interconnected contexts
• research baselines comparing graph retrieval to dense top-k and to other graph-RAG systems (GraphRAG, RAPTOR, LightRAG)