Large language models (LLMs) can greatly benefit from better integration of up-to-date information and reducing biases, which are often found in Retrieval-Augmented Generation (RAG) techniques. However, these models face challenges due to their complexity and longer response times. Therefore, optimizing the performance of RAG is key to their effectiveness in real-time applications where accuracy and timeliness are vital, such as in the field of medical diagnosis.
To address the current limitations of RAG models, methods involving workflows from query classification to summarization have been used. Query classification determines whether retrieval is necessary while retrieval methods such as BM25, Contriever, and LLM-Embedder work to obtain relevant data. Reranking then adjusts the order of the retrieved data and repacking organizes this data for better generation. Lastly, all key information is extracted in the summarization process. Despite these methods, problems remain, such as computation-intensive query rewriting and decomposition, and slow reranking with deep language models.
To counteract these problems, researchers from Fudan University conducted a detailed investigation of existing RAG methods and potential combinations to find optimal strategies. They carried out this research in three steps: comparing each RAG process, evaluating the impact of the method on the overall RAG performance, and studying promising combinations for various scenarios. The study suggests strategies that balance performance and efficiency. A significant contribution of the research is the incorporation of multimodal retrieval techniques into the RAG. This innovative method enhances question-answering capabilities when dealing with visual inputs and speeds up multimodal content generation.
The research went through several robust tests and evaluations to identify the most effective application for each RAG mechanism. The retrievals used two datasets, TREC DL 2019 and 2020, and methods such as BM25 for sparse retrieval and Contriever for dense retrieval. The team tested and improved the quality of the retrieval with different chunking sizes and techniques.
The outcome of the research resulted in substantial improvements across different performance metrics. Among all the different methods, the Hybrid with HyDE method achieved the highest scores on the TREC DL 2019 and 2020 datasets, with mean average precision values of 52.13 and 53.13 respectively. This is an excellent step over baseline methods.
In conclusion, the research successfully optimized RAG techniques to improve LLM performance. The researchers systematically evaluated existing methods, proposed innovative combinations of techniques, and demonstrated substantial improvements in performance metrics. The addition of multimodal retrieval techniques will prove to be a significant advancement in AI research. This research not only provides a robust structure for implementing RAG frameworks but also paves the way for future research in various domains to explore further optimizations and applications.
