Finding Connections: GraphRAG Workshop

Neo4j Nodes 2025 Conference Workshop

A hands-on workshop exploring GraphRAG — understanding not just what it is, but why it works and how each component contributes. The goal is to equip you with transferable skills and concepts you can remix for your own use cases.

🎯 Workshop Overview

GraphRAG combines knowledge graphs with retrieval-augmented generation to get better answers from language models. But GraphRAG is more than a single technique — it's an intersection of key technologies whose impact is still being discovered.

This workshop breaks down GraphRAG into its fundamental building blocks:

Topics Covered

Concept	What You'll Learn
Graphs	Nodes, edges, entities, and relationships
RAG	Retrieval Augmented Generation — putting relevant information into context
Semantic Search	Finding similar content through meaning, not just keywords
Embeddings	Converting text into numerical vectors that preserve semantic relationships
Cosine Similarity	Comparing vector directions in high-dimensional space
Structured Data Extraction	Getting structured output from unstructured text
Constrained Decoding	Guiding language models to produce schema-compliant responses
GraphRAG Pipeline	How all these pieces fit together end-to-end

🚀 Getting Started

Prerequisites

1. Create a Neo4j account: Visit console.neo4j.io and sign up
2. Create a free instance: Click "Create instance" in the Neo4j console
3. Save your credentials: Download the credentials file when prompted

Using the LLM Graph Builder

1. Visit llm-graph-builder.neo4jlabs.com
2. Click "Connect to Neo4j"
3. Upload your credentials file or enter connection details manually
4. Start exploring!

🔧 Key Technologies

Embeddings

Text is converted into numerical vectors using specialized embedding models. These vectors capture semantic meaning — similar concepts end up with similar directions in vector space.

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")

# "The sun is shining" → 384-dimensional vector
embedding = model.encode("The sun is shining")

Recommended model: all-MiniLM-L6-v2 — lightweight, fast, and effective for semantic similarity.

Structured Data Extraction

Language models can return structured data that fits a defined schema:

from pydantic import BaseModel
from openai import OpenAI

class CalendarEvent(BaseModel):
    name: str
    date: str
    participants: list[str]

response = client.responses.parse(
    input=[
        {"role": "system", "content": "Extract the event information."},
        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
    ],
    text_format=CalendarEvent,
)
# → CalendarEvent(name='Science Fair', date='Friday', participants=['Alice', 'Bob'])

This uses constrained decoding — the model is restricted to only generating tokens that satisfy the schema.

GraphRAG Pipeline

1. Preprocessing: Extract entities and relations from documents using structured data extraction
2. Embedding: Convert entities into vectors for semantic search
3. Storage: Add to a graph database (Neo4j)
4. Querying: Find similar entities via cosine similarity, traverse the graph for related information
5. Generation: Feed retrieved context to the language model alongside the question

📚 Resources

GraphRAG Tools

Tool	Description
Microsoft GraphRAG	Widely adopted, popular implementation
GraphRAG Visualizer	Visualize Microsoft GraphRAG outputs
nano-graphrag	Small, easy to hack on
Neo4j GraphRAG Python	Neo4j ecosystem integration
LLM Graph Builder	Interactive web tool for building graphs

Document Parsing

Language models work best with markdown. These tools help convert documents:

Tool	Description
Docling	IBM Research — document understanding
MarkItDown	Microsoft — document to markdown conversion

Embedding Models

• all-MiniLM-L6-v2 — 384 dimensions, fast and lightweight
• sentence-transformers — Framework for using and training embedding models

💡 Why This Matters

GraphRAG is a gateway technology. Each building block is valuable on its own:

• Embeddings power recommendation systems, anomaly detection, clustering, and more
• Structured extraction lets you build APIs over unstructured data
• Graph databases reveal hidden connections in any relational data
• RAG makes language models actually useful for domain-specific tasks

You don't need to adopt GraphRAG wholesale — understand the components, then remix them for your own problems.

👤 About the Presenter

Johan — AI/Data engineering consultant with over 15 years of building software, focused on ethical tech.

📂 Repository Contents

.
├── README.md          # This file
├── index.html         # Workshop presentation (Reveal.js)
├── CLAUDE.md          # Project context and guidelines
└── deepresearch.png   # Deep Research screenshot

📄 Slides

The presentation is built with Reveal.js and includes interactive visualizations:

• Graph visualization — D3.js force-directed graph showing entities and relationships
• Vector space visualization — 2D cosine similarity demonstration
• Token probability charts — How constrained decoding works
• Dimension distance distributions — Why cosine similarity beats Euclidean distance in high dimensions

To view the slides locally, simply open index.html in a browser.

---

Workshop presented at Neo4j Nodes 2025 — slides available at resolveworks.github.io/nodes2025