GraphRAG-Driven Agentic Reasoning: A Research Prototype using ADK, MCP, and Neo4j

Read full text Summarise with VIRAT
Authors: Jan Köster
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

GraphRAG-Driven Agentic Reasoning: A Research Prototype (v0.1.0) Summary: Initial Proof-of-Concept (PoC) release of a GraphRAG-driven reasoning system (case study: ancient numismatics), developed within the "Imagines Nummorum" project at the Berlin-Brandenburg Academy of Sciences and Humanities (BBAW). This release demonstrates the integration of structured research data as an external "Ground Truth" for LLM-based agents using the Model Context Protocol (MCP). Technical Features: Graph Architecture: Implements an experimental version of the IN.IDEA 4-layer model (Uni/Carrier, Formal Analysis, Epistemic, and Ontology layers) using Neo4j. Agentic Reasoning: Utilizes the Google Agent Development Kit (ADK) and Gemini 3.1 Flash Light to perform ad-hoc Cypher queries and morphological coin analysis. Multimodal RAG: Integration of CLIP-based vector search within a Knowledge Graph to bridge visual similarity and scholarly interpretations. Provenance & Uncertainty: Features reified nodes for interpretations and evidence to model scientific discourse transparently ("Glass Box AI"). Context: This prototype serves as a functional demonstration of how axiomatic, multi-layered graph schemas enable complex reasoning on small-scale LLMs while minimizing hallucinations and token consumption through topological Markdown outputs.

Fields of Study & Subjects

Digital Humanities and Scholarship Semantic Web and Ontologies Topic Modeling

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
23 March 2026
Language
EN