HippoRAG 2 - A search enhancement and generation framework developed by Ohio State University

What is HippoRAG 2?

HippoRAG 2, developed by Ohio State University, is a Retrieval Augmentation (RAG) framework that addresses the limitations of existing RAG systems in simulating the dynamics and relevance of human long-term memory. Based on a personalized PageRank algorithm, HippoRAG 2 integrates deep paragraphs and utilizes a more effective online LLM (Large Language Model) to push RAG systems closer to the performance of human long-term memory. In the offline phase, HippoRAG 2 uses LLM to extract triples from paragraphs and construct an open knowledge graph (KG), while also adding synonyms to the KG based on an embedding model. During online retrieval, the query is linked with triples and paragraphs in the KG, irrelevant information is filtered using LLM, and a personalized PageRank algorithm is applied for context-aware retrieval, ultimately providing the most relevant paragraphs for question-answering tasks.

Main features of HippoRAG 2

• Efficient knowledge retrieval and integrationBased on deep paragraph integration and knowledge graph (KG) construction, it can quickly retrieve and query relevant knowledge and integrate it into the generation process.
• Multi-hop association reasoningWith the help of a personalized PageRank algorithm, the system performs multi-hop reasoning, connects scattered knowledge fragments, and handles complex question-and-answer tasks.
• Context-aware retrievalBased on deep interaction between queries and knowledge graphs, the search results are dynamically adjusted according to the context, improving the accuracy and relevance of the search.
• Continuous learning abilityAs a non-parametric continuous learning framework, HippoRAG 2 can absorb and utilize new knowledge in real time without modifying model parameters, thereby enhancing the system's adaptability.

Technical Principles of HippoRAG 2

• Offline IndexingThe method uses LLM to extract structured triples (subject, relation, object) from text paragraphs and integrates these triples into an open knowledge graph (KG). Synonyms are detected using an embedding model, and synonym edges are added to the KG to enhance its connectivity. The original paragraphs are then combined with the knowledge graph to form a composite knowledge graph containing conceptual and contextual information.
• Online Retrieval:
  • Query linkThe query is matched with triples and paragraphs in the KG using an embedding model to determine the seed node for graph search.
  • Three-way filterThe retrieved triples are filtered based on LLM to remove irrelevant information and retain knowledge that is highly relevant to the query.
  • Personalized PageRank AlgorithmBased on the KG structure, a personalized PageRank algorithm is applied for context-aware retrieval, dynamically adjusting the relevance of retrieval results.
  • Paragraph ranking and Q&AParagraphs are ranked based on their PageRank scores, and the top-ranked paragraphs are used as context input into the final question-answering model.
• Personalized PageRank AlgorithmOne of the core technologies of HippoRAG 2 is a personalized PageRank algorithm that simulates the multi-hop reasoning process in human memory, performs deep search in the knowledge graph, connects scattered knowledge nodes, and better handles complex relational tasks.
• In-depth paragraph integrationThis approach deeply integrates paragraphs with nodes in the knowledge graph, preserving the contextual information of the paragraphs, enhancing the semantic richness of the knowledge graph, and making the search results more relevant and accurate.

HippoRAG 2 project address

• GitHub repository:https://github.com/OSU-NLP-Group/HippoRAG
• arXiv technical paper:https://arxiv.org/pdf/2502.14802

Application scenarios of HippoRAG 2

• Intelligent Question AnsweringQuickly answer complex questions and provide accurate answers.
• Knowledge ManagementIt enables efficient retrieval and recommendation of relevant content, improving the efficiency of knowledge utilization.
• Educational SupportReal-time updates of learning resources to support teaching and research.
• Medical consultationSearch for medical knowledge and provide comprehensive health advice.
• Law and FinanceRapidly integrate regulations and data to support expert decision-making.