Agora: A Distributed Language Model framework with API-call Support for Integrated Climate Forecasting

Abstract Large Language Models (LLMs) have achieved remarkable performance in a wide range of natural language processing (NLP) tasks and are increasingly applied to domain-specific challenges. However, integrating external knowledge into LLMs, thus allowing them to reason outside of their training data remains a significant hurdle. Additionally, the deployment of LLM-based systems demands substantial GPU resources, limiting the scalability of open-source solutions and driving a shift towards proprietary models, such as those from OpenAI and Amazon's Claude. In this paper, we present Agora, a system built on open-source language models for generating climate and agricultural forecasts and recommendations. Agora is designed to seamlessly integrate third-party API data, scale efficiently across diverse recommendation tasks, and operate on commodity GPUs without sacrificing performance. Moreover, Agora is trained to intelligently orchestrate API calls, enabling it to reason about data dependencies in user queries and dynamically determine the necessary API interactions.