A question can require evidence scattered across documents that share few words with it. A service-ownership document names a team, a git log identifies an engineer through a file path, and an on-call schedule records who covered an outage. A vector search may retrieve the first document without finding the chain needed to answer who last touched the service and whether they were on call. HippoRAG uses an entity graph to retrieve across those gaps. section 12.3
Offline, an LLM extracts subject–predicate–object triples from passages. Their noun phrases become graph nodes, with links back to passages; a dense encoder adds synonymy edges between equivalent phrases. The result is a relatively sparse associative map, not a set of detailed node descriptions or precomputed community summaries. section 12.3.2
At query time, retrieval follows four steps:
- Identify entities in the question as starting nodes, or seeds. section 12.3.2
- Run personalized PageRank from those seeds. Probability spreads along graph edges while periodically returning to the seeds, giving reachable nodes activation scores. section 12.3.2
- Aggregate node scores back to their linked passages, rank the passages, and give the top results to an LLM as answer context. section 12.3.3
This replaces an iterative retrieve–reason–retrieve loop with one graph-based retrieval pass. Unlike GraphRAG, HippoRAG does not prepare community summaries for questions about corpus-wide themes. It is a better fit when the task is to connect specific facts, the full summary pyramid is unnecessary, and a flat vector index cannot bridge the documents. section 12.3.3 section 12.4
Table 12.3 IRCoT vs HippoRAG pipelines
|
Step
|
IRCoT
|
HippoRAG
|
|
1
|
Retrieve top-k passages for the question
|
Extract entities from the question
|
|
2
|
LLM writes one reasoning sentence
|
Run PPR over the graph from those seeds
|
|
3
|
Use that sentence to retrieve more passages
|
Aggregate PPR scores back to passages
|
|
4
|
Repeat steps 2–3 until the answer is reached
|
Rank passages, hand the top ones to the LLM
|
|
LLM calls during retrieval
|
One per reasoning step (4–6 typical)
|
Zero
|
Compare the retrieval steps and LLM calls in iterative IRCoT and HippoRAG. section 12.3.3
A hub can hide a correct path
In an eleven-node supplier graph, a question about suppliers affected by a recall seeds HippoRAG with the named entities. Personalized PageRank reaches the correct supplier two hops away, but ranks it below nearer nodes around a well-connected hub. With only the top four nodes retrieved, the generator never sees it. Reaching a node is not the same as ranking it high enough to use. section 12.4.4
Evaluate retrieval against the questions that matter, especially paths passing through hubs and the cutoff used for context. HippoRAG trades community-wide summaries for cheaper indexing, but its graph construction still requires extraction from the corpus, and relevant multi-hop evidence can fall below the retrieval cutoff. section 12.3.2 section 12.4.4
