Version 1
: Received: 29 August 2024 / Approved: 29 August 2024 / Online: 30 August 2024 (03:45:59 CEST)
How to cite:
Yang, C.; Fujita, S. Adaptive Control of Retrieval-Augmented Generation for LLMs Through Reflective Tags. Preprints2024, 2024082152. https://doi.org/10.20944/preprints202408.2152.v1
Yang, C.; Fujita, S. Adaptive Control of Retrieval-Augmented Generation for LLMs Through Reflective Tags. Preprints 2024, 2024082152. https://doi.org/10.20944/preprints202408.2152.v1
Yang, C.; Fujita, S. Adaptive Control of Retrieval-Augmented Generation for LLMs Through Reflective Tags. Preprints2024, 2024082152. https://doi.org/10.20944/preprints202408.2152.v1
APA Style
Yang, C., & Fujita, S. (2024). Adaptive Control of Retrieval-Augmented Generation for LLMs Through Reflective Tags. Preprints. https://doi.org/10.20944/preprints202408.2152.v1
Chicago/Turabian Style
Yang, C. and Satoshi Fujita. 2024 "Adaptive Control of Retrieval-Augmented Generation for LLMs Through Reflective Tags" Preprints. https://doi.org/10.20944/preprints202408.2152.v1
Abstract
While Retrieval-Augmented Generation (RAG) enhances Large Language Models (LLMs), it also presents challenges that can affect model accuracy and performance. Practical applications show that RAG can mask the intrinsic capabilities of LLMs. Firstly, LLMs may become overly dependent on external retrieval, underutilizing their own knowledge and inference abilities, which can reduce responsiveness. Secondly, RAG techniques might introduce irrelevant or low-quality information, adding noise to the LLM. This can disrupt the normal generation process, leading to inefficient and low-quality content, especially when dealing with complex problems. This paper proposes a RAG framework that uses reflective tags to control retrieval. This framework evaluates retrieved documents in parallel and incorporates the Chain of Thought (CoT) technique for step-by-step content generation. The model selects the highest quality and most accurate content for final generation. The main contributions include: 1) Reducing the hallucination problem by selectively utilizing high-scoring document, 2) Enhancing real-time performance through timely external database retrieval, and 3) Minimizing negative impacts by filtering out irrelevant or unreliable information through parallel content generation and reflective tagging. These advancements aim to optimize the integration of retrieval mechanisms with LLMs, ensuring high-quality and reliable outputs.
Keywords
Retrieval-Augmented Generation; Large Language Models; Chain of Thought; reflective tag
Subject
Computer Science and Mathematics, Artificial Intelligence and Machine Learning
Copyright:
This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.