agentic-graph-rag

<p align="center"> </p>

Case Study: Ancient Numismatics & the IN.IDEA Model

DOI: [10.5281/zenodo.19183341](https://doi.org/10.5281/zenodo.19183341)

This prototype transforms a static Knowledge Graph into an interactive chat partner. Key capabilities include:

• Agentic Graph Traversal: The agent generates ad-hoc Cypher queries to navigate complex relationships without hardcoded paths.
• Epistemic Reasoning: Instead of just returning "facts," the system retrieves interpreted data, including scholarly confidence scores and evidence chains.
• Multimodal Analysis: Combines CLIP-based visual similarity with LLM-driven morphological analysis to identify and describe compositions.
• Context-Aware Onboarding: An MCP-driven onboarding process that injects the current graph ontology into the agent's context.
• Efficiency-First Architecture: Designed to interact with small-scale models (e.g., Gemini 3.1 Flash Light Preview with *Thinking: LOW*, or even *MINIMUM*). No expensive flagship models are required for this kind of reasoning - only high-quality, structured data (which is, admittedly, the hardest part to obtain).
• Token-Optimized Communication: Unlike many traditional RAG systems that flood the context window with verbose JSON, this system uses "topological" Markdown for detailed graph outputs (unit details) and GraphQL schema definitions. This drastically reduces token consumption while maintaining perfect legibility for the LLM.

The system consists of four main components orchestrated via Docker:

• Neo4j Database: Stores the multi-layered Knowledge Graph and handles vector-based similarity searches for coin images.
• **FMA