Can knowledge graphs enable multi-hop reasoning in one retrieval step?

Multi-hop question answering requires finding information that depends on other information: "Who was the mentor of the person who founded X?" requires finding X's founder, then finding that person's mentor. Standard RAG retrieves once per question. Iterative RAG retrieves multiple times, issuing new queries based on intermediate answers. Both approaches have costs: single-step RAG misses the chain; iterative RAG is slow and expensive.

HippoRAG takes a third approach, inspired by hippocampal indexing theory of human memory. The corpus is converted into a schemaless knowledge graph: LLMs extract entities and relationships from passages, forming a graph where nodes are concepts and edges are relationships. This happens offline, at indexing time.

At query time: extract key concepts from the query; seed Personalized PageRank (PPR) on the knowledge graph using those concepts. PPR propagates importance scores through the graph from the seed nodes, naturally following edges (relationships) and discovering connected subgraphs. Passages associated with highly-ranked graph regions are retrieved.

The multi-hop capability emerges naturally from graph traversal: PPR follows relationship edges, which means it discovers concepts multiple hops away from the query seeds. The traversal is complete in a single retrieval step because the graph encodes all multi-hop paths.

Results: 20% better than state-of-the-art on multi-hop QA; single-step HippoRAG matches or beats iterative retrieval (IRCoT) while being 10-20x cheaper and 6-13x faster. Integrating HippoRAG into IRCoT brings further gains.

The memory analogy holds architecturally: the knowledge graph functions as an associative index (analogous to hippocampus) over perceptual content (analogous to neocortex), enabling pattern-completion retrieval from partial query cues.

Complementary KG-based approaches: GraphRAG (Microsoft) uses knowledge graphs for a fundamentally different purpose than HippoRAG. Where HippoRAG uses the KG for traversal-based retrieval (find specific facts via graph navigation), GraphRAG uses KG + Leiden community detection for modular summarization — answering global corpus questions ("What are the main themes?") that no single retrieval step can address. The distinction: HippoRAG exploits graph connectivity; GraphRAG exploits graph modularity. See Can community detection enable RAG systems to answer global corpus questions?. SymAgent takes yet a third approach: deriving symbolic rules from KG structure to create navigational plans that align natural language with graph topology — using the graph not for retrieval or summarization but for structural reasoning guidance. See Can symbolic rules from knowledge graphs guide complex reasoning?.

Source: RAG; enriched from Knowledge Graphs