Note RAG do two things. Slicing & Retrieving. Slicing document into passages ignores the connection between passage, and the resulting "database" is not structured. Human stored their knowledge with more structure, related information of one concept are usually clustered together. HippoRAG adopts a LLM to process the knowledge base into a graph structure, continuously updating the knowledge graph with new information flow. During retrieving, most relevant nodes along with their neighbours are picked