Agricultural productivity in developing countries is increasingly imperiled by climate variability. Nepal presents a compelling case study: with 66% of its population dependent on agriculture, the country must contend with fragmented landholdings, erratic monsoon patterns, and severe agro-ecological heterogeneity across 1,359 distinct climatic zones. Timely, location-specific weather forecasts are a prerequisite for informed crop management, yet such systems have remained largely inaccessible to smallholder farmers. This study presents a comprehensive, data-driven agricultural decision-support framework that integrates spatio-temporal weather forecasting, soil-aware crop recommendation, and a Retrieval-Augmented Generation (RAG) advisory chatbot within a unified mobile platform, with the aim of enhancing farm-level decision-making across Nepal's diverse agro-ecological landscape. Historical weather data spanning 42 years from 1,359 locations (NASA POWER) were used to train two deep learning architectures: a Transformer-based Graph Neural Network and a Spatio-Temporal Graph Convolutional Network (STGCN). A graph structure encoding geodesic distance and altitudinal similarity (edges within 15 km or 50 m altitude difference) modeled spatial dependencies. A 120-day rolling lookback window captured temporal dynamics. Forecasted weather variables were integrated with static soil attributes via an inverse-distance scoring algorithm to rank crop suitability. A RAG chatbot using hybrid SPLADE and dense embeddings provided natural-language advisory support in Nepali and English. The STGCN achieved superior forecasting performance with a Mean Squared Error (MSE) of 0.011 compared to 0.013 for the Transformer-based model, demonstrating its capacity to capture complex spatial and temporal dependencies. The crop recommendation engine generated ranked suitability indices across all 1,359 locations. The RAG-based advisory system produced contextually relevant, multilingual responses to diverse farmer queries. The mobile deployment received positive qualitative feedback from rural users in terms of accessibility and relevance. This study demonstrates that integrating spatio-temporal deep learning with soil data fusion and conversational AI can deliver scalable, accessible, and accurate agricultural guidance in data-scarce, geographically complex settings. The framework offers a replicable model for precision agriculture in other developing-country contexts vulnerable to climate variability.