The current state of data wealth, information scarcity paradox in is Modern Controlled Environment Agriculture. Modern networks of Internet of Things (IoT) sensors produce enormous amounts of untapped streams of telemetry (data) in real-time, the data is framed to the farmers via dashboard interfaces, which impose large cognitive overhead to the farmer by requiring that (s)he correlate the numeric and graphic data points, manually, into actionable insights. Simultaneously, Large Language Models (LLMs) can provide wrappers of impressive reasoning and natural language interaction capabilities, but are ultimately decoupled from the physical world, and thus their advice is generic and lacks situational awareness. We present a novel framework that fills in this important gap in the literature. Here, we propose an architecture that integrates conversational AI with real-time IoT data using a Dual-Source Retrieval-Augmented Generation (RAG) pipeline. This pipeline makes the LLM's responses grounded by dynamically fetching context from two independent sources. Using a sensor network of ESP32 sensors combined with a cloud-based LLM, the system enables farmers to query the system using natural, spoken queries and use scientifically-sound, contextually relevant decision support. We validate our prototype by showing low end-to-end latency (4.21 s mean) and a large qualitative improvement in both response actionability and safety compared to non-grounded models.