Masterarbeit
Optimizing SLMs for tool-calling through context and prompt management
Research Area
Intelligent Information Management
Advisers
Lucas Schröder
Description
Small Language Models (SLMs) are becoming increasingly capable, enabling the use of AI models directly on consumer or edge hardware. However, they still fall short compared to larger models, often due to a more limited context size or accuracy. This is especially relevant for the topic of local Web Agents, which rely on accurate tool calls for interacting with web applications, and where available tools change frequently depending on the visited web applications. The goal of this thesis is to investigate possible optimizations of SLMs for tool calling through context and prompt management approaches, especially for frequently changing sets of tools, including the implementation and evaluation of a suitable prototype.
