Imagine an AI that could sift through millions of patient records, instantly identifying individuals with rare diseases and revealing hidden connections between symptoms and treatments. This isn't science fiction—it's the promise of knowledge graphs powered by large language models (LLMs). Researchers are exploring how LLMs can extract structured medical knowledge from the messy, unstructured text of electronic health records. This structured data is then organized into knowledge graphs, creating a powerful tool for summarizing patient information and enabling faster, more accurate searches for specific conditions. The challenge lies in teaching AI to understand the nuances of medical language and connect it to standardized medical ontologies. Early experiments show that while LLMs excel at processing text, they can benefit from techniques like 'gleaning' and 'few-shot learning' to improve accuracy. Gleaning involves repeatedly refining the AI's extraction process, while few-shot learning provides the AI with specific examples to guide its understanding. Researchers tested their approach on a real-world dataset of over 33 million patients, focusing on rare diseases like Dravet syndrome and BPAN. For Dravet syndrome, where diagnostic codes exist, the AI successfully identified patients and extracted key phenotypic features from their medical histories. This allowed researchers to compare real-world symptom frequencies with those documented in the Human Phenotype Ontology (HPO), a standardized vocabulary of human phenotypes. The results highlighted important discrepancies, suggesting the need for more dynamic, data-driven approaches to disease characterization. For BPAN, which lacks specific diagnostic codes, the challenge was even greater. The AI had to search for patients based on broader, less precise codes and then analyze their medical notes for phenotypic features associated with BPAN. This process identified a small group of potential BPAN cases, highlighting the AI’s potential to uncover hidden patient populations. While promising, this technology is still under development. Challenges remain in ensuring accuracy and addressing ethical considerations around patient privacy. However, the potential of LLM-powered knowledge graphs to revolutionize medical research and diagnosis is undeniable. By unlocking the wealth of information hidden within electronic health records, AI could pave the way for faster diagnoses, more personalized treatments, and a deeper understanding of rare and complex diseases.