Retrieval and Reasoning on Graphs

Graphs naturally model many real-world data and power a variety of industry use cases, such as fraud detection in financial networks, friend recommendation in social networks, and route planning for logistics. The technologies involved include graph databases and query languages, algorithms, and machine learning models. In recent years, frontier LLMs have shown remarkable natural language understanding and reasoning capabilities. Retrieval-augmented generation (RAG) methods allow LLMs to access private data, and recent LLM-based agentic systems are beginning to solve more complex tasks, such as coding, that involve collaborating with users. Promising directions are emerging to make LLMs work effectively on graph data.

In this talk, we will present the latest Neo4j developments in this space. Specifically, we will demo our GraphRAG framework and an agent that can call graph algorithms for reasoning. This enables users to ask any question in natural language about their graphs and receive answers. We will dive into the engineering and research challenges, present our benchmarks and findings, and discuss future work.

Brian Shi is a senior software engineer at Neo4j, where he leads the work on graph machine learning and algorithms. More recently, he focuses on using techniques and tools from graph ML and algorithms to make LLMs work better on graphs.