Interaction information measures how much one passage’s information about an answer changes after another passage is known. For answer Y and passages X1 and X2, it is I(Y;X1) − I(Y;X1|X2). section 12.5
Read the sign as a practical retrieval diagnostic:
- Positive values indicate redundancy: X2 makes X1 less informative because both passages communicate overlapping information. section 12.5
- Negative values indicate synergy: X1 becomes more informative after X2 is known, so the passages answer the question together. section 12.5
- This distinguishes graph retrieval from vocabulary-driven vector retrieval: cosine similarity can select redundant chunks, while graph traversal aims to connect complementary evidence. section 12.5 section 12.7
Evaluating a retriever ensemble
Suppose one passage describes a regulation and another identifies the affected entity, while neither answers the query alone. Their negative interaction information indicates the desired multi-hop synergy. Compare that score with marginal contribution, Multi-hop Success Rate, Hallucination Rate, and Graph Completeness before adding graph infrastructure. section 12.5 section 12.7
Table 12.4 Graph-based RAG comparison
|
Approach
|
Taxonomy
|
Core idea
|
Reasoning power
|
Cost (build/query)
|
Best use case
|
|
Microsoft GraphRAG
|
Global summarization
|
Pre-computes summaries for hierarchical communities
(Leiden clusters)
|
Very high
|
High
|
Global summarization ("sense-making")
|
|
HippoRAG
|
Associative memory
|
Uses Personalized PageRank (PPR) to spread activation to
hidden links
|
High
|
Moderate
|
Discovery (scientific research)
|
|
LightRAG
|
Hybrid (text + graph)
|
Uses graph structure but retrieves via efficient vector
search (dual-path)
|
Medium
|
Low
|
Production enterprise RAG (real-time)
|
|
GRAG (Hu et al.)
|
KG-first
|
Explicit knowledge-graph traversal linearised into LLM
context
|
High
|
High
|
Deep reasoning (compliance, legal)
|
Compare graph-based RAG architectures by reasoning power, construction and query cost, and intended use when choosing a retrieval design. section 12.5
