Abstract
This study investigates how Retrieval-Augmented Generation (RAG) systems can be designed to support policy analysis, with a focus on identifying and analysing links between audit recommendations and subsequent government measures in Swedish policy documents. The study designs, implements and evaluates a task-specific RAG system tailored to the analysis of three types of Swedish policy documents: government communications, committee terms of reference and budget bills. The system combines semantic retrieval with Large Language Model (LLM)-based analysis and structured outputs to support systematic analysis of policy measures. A central focus of the study is the role of chunking strategies in RAG systems, comparing fixed token-based, fixed character-based and recursive character-based strategies. The evaluation combines automatic label-based methods with rubric-based manual assessment to compare RAG system performance across chunking strategies and document types. The results show that the system overall performs well and produces stable, well-grounded outputs, indicating that relevant contextual information is effectively retrieved and utilised within the system. While differences between chunking strategies are observed, no single strategy consistently outperforms the others. Instead, the results suggest that performance is context-dependent, varying across document characteristics and analytical tasks. Recursive chunking performs well for the more structured documents included in the study, whereas fixed-size strategies are more effective for longer and more complex documents, such as budget bills. Overall, the study demonstrates that RAG systems can support systematic analysis of Swedish policy documents and contribute to partially automating complex analytical tasks. It further highlights the importance of aligning system design choices with both document characteristics and the analytical task at hand.
Showing the abstract — retrieve the full paper via the Exa API.