🧠 LLMs meet Knowledge Graphs: Exploring RDF graphs using Neuro-Symbolic AI

TU/e Built Environment Coding Cafe

Alex Donkers

Quick links → Gemini API key → Notebook

This repository contains all materials for the "LLMs meet Knowledge Graphs" workshop — a hands-on introduction to combining RDF knowledge graphs, ontologies, and Large Language Models (LLMs) for reasoning over structured data.

Participants will learn how to go from raw regulation text to a queryable, explainable knowledge system, powered by FireBIM ontologies and Gemini 2.0.

---

🎯 Learning Goals

By the end of this workshop, you will be able to:

• Load and explore RDF data in Python.
• Understand and query ontologies with SPARQL.
• Use LLMs to classify text and enrich knowledge graphs.
• Translate natural language questions into SPARQL queries.
• Build a neuro-symbolic GraphRAG pipeline that connects LLM reasoning to symbolic RDF structures.

---

🧩 Workshop Modules

Module	Topic	Description
1. GraphDB & RDF Exploration	Understanding graphs	Explore RDF data using rdflib and visualize it interactively with PyVis.
2. RDF & OWL Ontologies	Reasoning over structure	Load both the instance graph and the FireBIM Regulation Ontology (FRO) to understand how semantics and relationships are defined.
3. Large Language Models (LLMs)	Neural meets symbolic	Use Gemini 2.0 Flash to classify regulation texts into high-level themes and enrich the RDF graph with semantic triples.
4. GraphRAG & Natural Language Querying	Neuro-symbolic reasoning	Generate SPARQL queries from plain English questions and summarize results using an LLM.

---

🧱 Dataset

We use a small curated dataset of digital building regulations in Turtle (.ttl) format:

• fro.ttl – The FireBIM Regulation Ontology, defining classes and relationships.
• DutchFireRegulations.ttl – Instance data containing building regulation clauses.

---

🚀 Getting Started

1. Generate a Gemini API key → Google AI Studio

2. Open the notebook in Google Colab
   → Notebook

3. Run all cells sequentially
   Each module builds on the previous one — make sure you complete them in order.

4. Follow along the presentation (20 min)
   The slides summarize the theory and connect to each notebook exercise.

---

🧠 Mini Challenges

Each notebook includes hands-on challenges such as:

• Writing SPARQL queries to explore predicates and classes.
• Linking clauses to ontology concepts.
• Using Gemini to classify rules by topic.
• Translating natural language to SPARQL queries.
• Summarizing query results with Gemini.

Try them all to deepen your understanding!

---

🧩 Tech Stack

• 🧮 RDF / SPARQL: rdflib
• 🔍 Visualization: networkx, pyvis
• 🤖 LLMs: Gemini 2.0 Flash API
• 🧱 Ontology: FireBIM FRO and FBO
• 📓 Environment: Google Colab (no local GraphDB required)

---

🧠 About the Approach

This workshop demonstrates neuro-symbolic AI — combining:

• Symbolic reasoning: via RDF, OWL, and SPARQL
• Neural reasoning: via Gemini’s natural language understanding

Together, they enable GraphRAG (Graph Retrieval-Augmented Generation):

Natural language → LLM → SPARQL → Graph → Reasoned answer.

---

🧑‍🏫 Author

Alex Donkers
Built as part of the FireBIM research initiative on digital building regulations.
📧 Personal research page
🌐 LinkedIn

---

📜 License

Reuse of the material in this library is limited, and can only happen after personal agreement from the author.

---

⭐ If you use or adapt this workshop

Please cite it or link back to the GitHub repository to support open, explainable AI for the built environment.