This work presents a real-time context-awareness frame work for mobile robots navigating dynamic indoor environments. The system provides reliable semantic information that is mapped onto Risk Priority Numbers (RPNs) using Failure Mode and Effects Analysis (FMEA), enabling risk-aware navigation decisions that prioritise both safety and efficiency. The pipeline combines RGB-D perception with large language model (LLM)-based classification, assigning each detected object a subtype and context-aware attributes such as animacy, movability, obstruction potential, and attentiveness. These attributes are dynamically converted into RPN scores, guiding the robot to proceed, slow down, reroute, or stop based on the current scene. A modular ROS 2 architecture manages real-time data fusion, including a Context Conduit Node (CCN) that synchronises semantic outputs, fallback detections, and depth data into a unified, time-aligned world state. The framework incorporates a fallback mechanism that ensures continuous operation when LLM inference is delayed. Evaluation in simulation demonstrates that unified prompting improves attribute consistency and risk estimation compared to split prompting strategies, supporting responsive and risk-aware navigation decisions.