A Multi-Agent LLM Pipeline for Complex Report Generation from Multiple Sources with Human-in-the-Loop Feedback

Sammanfattning

Municipal waste management (MWM) plays an important role in preserving good environmental quality for urban populations. To support decision-making within MWM, large volumes of reports must be produced, covering a wide range of domains. The creation of these complex reports often requires manually checking data from several sources, a highly time-consuming task that can potentially delay decision-making. This thesis proposes the implementation of a multi-agent solution for generating reports, providing a semi-autonomous system that promotes work efficiency and reduces manual labor, thereby allowing professionals to allocate more time to strategic and operational planning. More precisely, we investigate how to design a pipeline that can automate the generation of reports using both structured and unstructured data sources. Additionally, we evaluate to what extent the provided pipeline improves report generation in terms of work efficiency, usability, and report quality. By combining Design Science Research and the Technology Transfer Model, we conducted a research and development process well suited for a company-academia collaboration. The results show that the system to some extent integrates structured and unstructured data and supports modular report generation. While users found the interface intuitive and the architecture promising, experts noted weaknesses in reproducing professional tone and company-specific style. The feedback loop was inconsistent, and no clear workload reduction was observed. These findings confirm the potential of the architecture, but highlight the need for additional optimizations, such as a better style extraction model and a refined prompt design.