✨ New course! Enroll in Agentic Knowledge Graph Construction

[Community-Team]

Enroll Now!

Learn to automate the construction of knowledge graphs using agents in Agentic Knowledge Graph Construction, taught by Andreas Kollegger, Developer Evangelist for Generative AI at Neo4j.

In traditional RAG systems, documents are split into chunks stored in a vector database. In a knowledge graph, chunks are additionally placed in a graph that better represents relationships within your data. For example, a chunk representing a product review can be connected in a graph to another node representing the product that was mentioned in the review. Manually constructing knowledge graphs can be a lot of work. In this course, you’ll learn how to use collaborative agents to generate the construction plan for your knowledge graph.

You’ll implement an agentic system using Google’s Agent Development Kit (ADK), to build a knowledge graph that helps you find the root cause of product issues. You’ll work with structured data consisting of product and supplier information, and unstructured data consisting of product reviews.

You’ll design agents that suggest how to transform your structured and unstructured data into graphs. For example, from each CSV file, you can either extract a node representing a product, a part of a product or a supplier, or you can extract a relationship: a product contains this part, a part is provided by this supplier. From each review chunk, you can extract what product and issues were mentioned. Finally, you will construct the graphs based on the plans provided by the agents and connect them in a complete knowledge graph.

[Maxli25]

This is an excellent instructional course covering knowledge graph design and several design patterns in agent development, using Google ADK. In the structured data processing chapter, I used the Qwen3:32B model for execution but consistently failed to achieve the graph schema output results as demonstrated in the video. Consequently, I adjusted various prompts and attempted to make the model meet the output format requirements. However, I noticed that initially, the model’s reasoning followed a certain logic and approach: it could retrieve the user’s goal, confirmed files, sample the files, consider node and relationship generation, and invoke propose_node_construction and propose_relationship_construction. Yet, each time it did not call get_proposed_construction_plan to output the generated node and relationship schema; instead, it directly returned the steps. I would like to ask about optimization directions: Should I first let the model generate a to-do list and then record each step until the final one is completed?