Utvärdering av avancerade tekniker för hämtningsförstärkt generering vid multi-hop frågor : En jämförande studie av Naive RAG, Recursive RAG och Graph RAG

Retrieval-Augmented Generation (RAG) enhances large language models (LLMs) by incorporating information retrieved from external sources, such as document corpora or structured databases, to improve the accuracy and relevance of generated responses. RAG techniques range from simple, naive approaches to advanced methods using recursive retrieval or structured knowledge. While naive RAG handles basic queries well, it struggles with multi-hop QA and is prone to hallucinations. More sophisticated methods improve reasoning but typically require greater computational resources. To assess the trade-off between accuracy and resource consumption, this thesis presents a comparative evaluation of four RAG systems: Naive RAG, Recursive RAG, KG-Graph RAG, and Community-Graph RAG. The evaluation is conducted on a multi-hop QA dataset with diverse query types, including inference, temporal, comparison, and null queries. Each system is assessed based on generation accuracy, retrieval quality, and resource consumption. KG-Graph RAG achieves high accuracy, strong inference capabilities, and the lowest hallucination rate, while maintaining efficient token usage and moderate computational demand. Community-Graph RAG excels in temporal queries but incurs higher latency and cost. Naive RAG and Recursive RAG offer a balanced trade-off between cost and accuracy, making them suitable for general-purpose applications. While these results offer practical guidance for selecting RAG strategies based on use-case needs, accuracy on multi-hop QA remains an area for improvement. Future work could explore hybrid systems that dynamically switch retrieval strategies based on query type to optimize both performance and efficiency.