Type-aware graph rankings offer a way to guide AI retrieval

What the study asks

Rosario Napoli, Antonio Celesti, Massimo Villari and Maria Fazio propose a system called Typed-Centrality Grounding for Agentic GraphRAG. Their paper was posted to arXiv on November 6, 2026. It is a preprint, not established peer-reviewed evidence.

Graph retrieval-augmented generation, or GraphRAG, supplies a language model with information from a knowledge graph: a network of entities and their relationships. The authors investigate whether graph structure can provide additional guidance beyond textual relevance.

Their focus is centrality, a family of measures describing a node’s structural importance. A node might rank highly because it has many connections, links otherwise separated parts of a network, or connects to other highly ranked nodes. The question is whether rankings across an entire mixed network obscure importance within particular entity types.

How the proposed system works

The framework has four layers: graph storage, a symbolic layer that calculates structural measures, a retrieval layer, and an agent that coordinates queries and generates responses.

The symbolic layer builds type-specific subgraphs and calculates five measures: degree, betweenness, closeness, PageRank and eigenvector centrality. These capture different aspects of connectivity, bridging, distance and influence.

Where entities of the same type have direct relationships, those connections provide the subgraph. Where they do not, the system connects same-type entities through shared intermediate neighbors. This operation, called a one-mode projection, can reveal associations absent as direct links in the original graph. Isolated entities are excluded from centrality calculations.

The resulting scores accompany textual information during retrieval and reasoning. They are deterministic calculations for the chosen graph representation, rather than additional model-generated descriptions. However, that does not make the language model’s final answer deterministic or guarantee its correctness.

What the researchers tested

The experiments use a POLE—Person, Object, Location, Event—dataset derived from open crime data for Greater Manchester. Its graph includes people, crimes, locations, postcodes, phones and other entities. The implementation uses Neo4j, with agent orchestration through Claude Desktop and the Model Context Protocol.

The paper compares configurations using global centrality measures and the proposed symbolic layer. It presents two investigative-analysis prompts, then examines the graph scores associated with the selected answers.

In the first demonstration, the agent was asked to identify the postcode acting as the most critical structural bridge in the crime network. The global-metric baseline selected WA14 2SP, with reported global betweenness of 38.5. The typed configuration selected M1 3EZ, with typed degree of 114.

The authors attribute the baseline’s choice to connections outside the relevant crime context. For M1 3EZ, they report 288 distinct paths connecting 30 crime events across seven categories, offering structural support for its selection within that context.

In the second demonstration, the agent was asked to identify the most influential crime node based on its connections. The baseline selected node 57116, with global betweenness of 68.75. The typed configuration selected cc45, whose reported PageRank changed from 0.2775 globally to 0.9612 in the typed projection, while degree changed from 2 to 49.

These results show that changing the graph representation and supplying type-specific metrics changed the agent’s selections. The authors interpret those changes as better alignment with the relevant subgraphs.

What the evidence does not establish

The evaluation is narrow: one dataset and two reported prompts. It provides no aggregate answer-accuracy score, measured hallucination rate, repeated-run variability or statistical uncertainty. The reported comparisons therefore do not quantify a general reliability improvement.

The mathematical argument also needs qualification. It compares average scores in a normalized global graph and a smaller typed graph, producing a ratio based on their node counts. But the paper’s own definition of degree is a connection count, not a probability distribution summing to one. An increase in average normalized scores does not, by itself, establish that an individual node is ranked more correctly.

Projection also changes the connections being measured. A higher projected degree or PageRank is not automatically evidence of better retrieval. The demonstrations compare selections supported by different centrality measures, including global betweenness and typed degree, rather than isolating a single metric under otherwise identical conditions.

The contribution is an interpretable way to supply type-specific structural signals to an AI agent. Its broader effects on accuracy, robustness and reasoning remain unestablished by the reported experiments. In this dataset, structural importance describes connectivity—not criminal responsibility or investigative significance beyond what those connections support.