TTFT-Aware Graph Chain-of-Thought:Distance-Indexed Neural A* for Low-Hallucination Multi-Hop Medical Reasoning

The article introduces a GraphRAG stack designed to enhance multi-hop medical reasoning in clinical LLMs, using a ~700K-node medical knowledge graph. It employs a directed Pruned Landmark Labeling (PLL) oracle for efficient distance checks and a lightweight AStarNet heuristic to prioritize clinically relevant paths, resulting in improved Time to First Token (TTFT) and reduced hallucinations during fertility-focused queries. This approach offers a practical solution for building explainable AI systems in healthcare, addressing critical issues of hallucination and reasoning transparency.