Knowledge Graph Extraction from Retrieval-Augmented Generator: An Application in Aluminium Die Casting

We present a novel, efficient, and scalable approach for generating knowledge graphs (KGs) tailored to specific competency questions, leveraging large language model (LLM)-based retrieval-augmented generation (RAG) as a source of high-quality text data. Our method utilises a predefined ontology and defines two agents: The first agent extracts entities and triplets from the text corpus maintained by the RAG, while the second agent merges similar entities based on labels and descriptions, using embedding functions and LLM reasoning. This approach does not require fine-tuning or additional AI training, and relies solely on off-the-shelf technologies. Additionally, due to the use of RAG, the method can be used with a text corpus of arbitrary size. We applied our method to the high-pressure die casting domain, focusing on defects and their causes. In the absence of annotated datasets, manual evaluation of the resulting KGs showed over 90% precision in entity extraction and around 70% precision in triplet extraction, the main source of error being the RAG itself. Our findings suggest that this method can significantly aid in the rapid generation of customised KGs for specific applications.