Retrieval Augmented Generation Optimizations

Abstract

This product-based thesis explores the optimization of Retrieval-Augmented Generation (RAG) systems, focusing on improving the precision and relevance of skill and occupation retrieval from CV texts. The study aims to enhance an existing Proof of Concept by incorporating advanced RAG techniques. Key areas of improvement include optimizing the RAG workflow, enhancing user interfaces for CV uploads, and integrating more advanced embeddings models and vector databases like FAISS. Various RAG techniques, including Hypothetical Document Embedding (HyDe), query decomposition, and hybrid search, were evaluated for their ability to enhance retrieval accuracy. While some improvements were observed, particularly in skills retrieval, the results showed limited overall gains. This thesis also proposes future research directions such as leveraging ColBERT embeddings and investigating longer context retrieval for large language models. The new Proof of Concept system provides a foundation for future development in career path guidance and job-matching features, with several recommendations for backend, frontend, and data security improvements.

The development and results of this work can be found here: https://github.com/robinrolle/CareerBot-RAG