Integration of External APIs into LLM-Based Agents for Enhanced Functionality: Comparing a Price Estimator to a Products Deal Finder
Year: 2025
Type: dissertation
Abstract: Large-language models (LLMs) have progressed from research curiosities to multi-tasking fully autonomous AI agents, but their reasoning is still restricted by frozen parameters and the text in their prompt pane. This thesis investigates how the controlled exposure of LLMs to external application-programming interfaces (APIs) may bypass that restriction and overall increase the scope of field applications for these technologies. To accomplish this task, investigate and curate an Amazon product listings dataset, testing it in four different LLM-based configurations. From a zero-shot GPT-4o-mini baseline to a multi-agent Deal Finder that orchestrates both a fine-tuned GPT-4o-mini model, a RAG model, RSS scrapers, ensemble regression and a Pushover push-notification API. The results confirm that the targeted API integration, even light-weight ones, significantly enhance LLM decision making and action capabilities. They moreover uncover new engineering trade-offs like rate limits, quality of retrieval, ensemble coordination, leaving to future work the possibility to scale API-centric agents from prototypes to production systems.
Author: Machado, João Duarte Justo Mirotes de Almeida
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7112895579