Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

Each to Their Own: Exploring the Optimal Embedding in RAG

Created by
  • Haebom

Author

Shiting Chen, Zijian Zhao, Jinsong Chen

Outline

This paper proposes two novel approaches to address the Limitations problem of Retrieval-Augmented Generation (RAG), which has attracted attention as a method for integrating recent information into large-scale language models (LLMs) or building domain-specific models. RAG utilizes multiple embedding models, but their heterogeneous characteristics lead to discrepancies in the similarity calculation results and the quality of the LLM responses. To address this issue, we propose Mixture-Embedding RAG and Confident RAG. Mixture-Embedding RAG integrates the retrieval results of multiple embedding models but fails to improve performance over conventional RAG. On the other hand, Confident RAG generates responses multiple times using multiple embedding models and selects the response with the highest confidence. This approach improves performance by approximately 10% and 5% over conventional LLM and RAG, respectively. Consistent results across various LLMs and embedding models demonstrate that Confident RAG is an efficient plug-and-play approach applicable to a wide range of fields.

Takeaways, Limitations

•
Takeaways:
◦
Confident RAG presents an efficient method to improve the performance of RAG by effectively combining the strengths of various embedding models.
◦
It shows consistent performance improvements across various LLM and embedding models, suggesting broad applicability.
◦
It is implemented in a plug-and-play manner, so it can be easily integrated into existing systems.
•
Limitations:
◦
Mixture-Embedding RAG did not show any performance improvement over conventional RAG. Further research is needed to improve it.
◦
There is a lack of detailed information about how Confident RAG measures reliability. There is room for improvement in how it measures reliability.
◦
Additional experiments across various domains are needed. The current results may be limited to specific domains.
👍
Made with Slashpage