eidos: A modular approach to external function integration in LLMs

Function calling allows Large Language Models (LLMs) to execute a wide range of tasks, from data analysis and mathematical computations to interacting with web services and other software systems. By harnessing the power of external tooling, LLMs can provide more dynamic, context-aware responses. However, errors in the model’s understanding of the request can lead to misinterpretations of the intended actions, resulting in function calls that are either irrelevant or incorrect for the task at hand. Without proper validation and control mechanisms, the parameters expected by the function may not align with those provided by the model, leading to incorrect operations or failures in task execution. In this paper, we present eidos, a software tool designed to streamline the integration of functions within LLMs. eidos acts as an intermediary, enabling both the seamless execution and validation of functions by LLMs. By leveraging its modular architecture, function definitions can be injected into the LLM context and invoked as if they were native functions via an API.