A system architecture that pairs an unsupervised, self-updating knowledge graph with a fine-tuned large language model so the KG acts as a Retrieval-Augmented-Generation guardrail: the verified graph supplies grounded context that prevents the LLM from hallucinating and keeps answers current past the model's training cut-off. Key ideas include continuously ingesting new peer-reviewed literature into a triplestore, classifying and clustering domain tables with a Bi-GRU over BioBERT embeddings, a structural search engine that separately indexes horizontal and vertical metadata of complex non-first-normal-form tables, and embedding-based schema matching used to fuse newly extracted subtrees into the existing graph. Instantiated as CancerKG.ORG for colorectal cancer but designed to be domain-independent.
An applied computer-science systems paper (cs.IR/cs.AI) describing CancerKG.ORG, a Web-scale hybrid of a Knowledge Graph (KG) and a Large Language Model (LLM) for retrieving verified, up-to-date know…