GraphRAG PhD Course Project: Practical Integration of Graph Neural Networks and Retrieval-Augmented Generation

This project is a teaching practice combining Graph Neural Networks (GNN) and Retrieval-Augmented Generation (RAG). Based on the Lord of the Rings dataset, it demonstrates how structured subgraphs enhance the reasoning ability of Large Language Models (LLMs). The project provides a directly runnable Dockerized application, aiming to explore the core concepts and technical implementation of GraphRAG, suitable for teaching and research scenarios.

Traditional RAG technology relies on text chunk retrieval, but isolated fragments lack structured relationship information, making it difficult to handle complex entity relationship reasoning problems. This project raises a core question: How does the reasoning ability of LLMs change when their context shifts from plain text to structured subgraphs? By building a complete GraphRAG demonstration system, the project shows the enhancement effect of graph structures on LLM reasoning, and the system supports Dockerized local deployment.

The project integrates two data sources to build a hybrid knowledge graph:

1. Raphtory LOTR Interaction Graph: A character interaction network based on sentence co-occurrence, containing rich structural information suitable for GNN input;
2. LOTRO OWL Ontology: In RDF/OWL format, including semantic relationships such as friendOf and hasWeapon for classes, characters, locations, etc. The hybrid structure retains narrative co-occurrence relationships (INTERACTS_WITH) and semantic type relationships, and adds attribute indicators like race, PageRank, and community affiliation.

The project uses a fully localized tech stack with no external API dependencies:

• Docker Compose: Container orchestration;
• Neo4j Community: Graph database storage and querying;
• FastAPI: Backend API framework;
• Ollama: Local LLM inference (supports default model qwen3.6:latest, gemma4:26b, etc.);
• Python: Data import, indicator calculation, and retrieval logic. The localized design is suitable for teaching demonstrations and privacy-sensitive scenarios.

The core functions of the project include:

1. Entity Detection and Subgraph Retrieval: After identifying entities in the question, retrieve k-hop neighborhood subgraphs in Neo4j as structured context;
2. Hop Count Comparison: Support adjusting hops (1/2/3) to demonstrate the impact of structural information in different scopes on answers;
3. Visual Exploration: Run Cypher queries via Neo4j Browser (e.g., Frodo's neighborhood, characters with the highest PageRank, shortest path between Frodo and Sauron).

The project shows the deep connection between GraphRAG and GNN:

• k-hop neighborhoods correspond to the receptive field of GNN (the range affecting node representation);
• Implicit aggregation when LLMs read subgraphs is analogous to the message-passing mechanism of GNN;
• PageRank/community affiliation as structural features are similar to GNN node features; Both utilize graph structure information, but GNN encodes graph embeddings through parameterized learning, while GraphRAG relies on LLM pre-trained knowledge to directly understand graph structures.

The project has the following limitations:

• The INTERACTS_WITH relationship is based on sentence co-occurrence, which does not equal causal or direct interaction;
• The LOTRO ontology is small in scale, and practical applications require support from larger ontologies. Expansion directions include: exploring the sufficiency of k-hop neighborhoods, applicable scenarios of Graph RAG vs. Vector RAG, limitations of co-occurrence relationships, input value of knowledge graph semantic features for GNN, how GNN optimizes GraphRAG subgraph ranking, etc.

As a PhD course tool, the project has significant teaching and practical value:

• Clear teaching process: Covers links such as motivation introduction, graph model introduction, Neo4j visualization, GraphRAG demonstration, GNN correlation, limitations and expansion;
• Reasonable technology selection: Neo4j facilitates visualization and querying, Ollama local models avoid external dependencies, Docker ensures reproducibility;
• Open-source and extensible: Provides clear research questions and expansion directions, suitable for researchers and students in the cross-field of GraphRAG and GNN.