Auto-RAG: Autonomous Retrieval-Augmented Generation for Large Language Models

Abstract: Iterative retrieval refers to the process in which the model continuously queries the retriever during generation to enhance the relevance of the retrieved knowledge, thereby improving the performance of Retrieval-Augmented Generation (RAG). Existing work typically employs few-shot prompting or manually constructed rules to implement iterative retrieval. This introduces additional inference overhead and overlooks the remarkable reasoning capabilities of LLMs. In this paper, we introduce Auto-RAG, an autonomous iterative retrieval model centered on the LLM's powerful decision-making capabilities. Auto-RAG engages in multi-turn dialogues with the retriever, systematically planning retrievals and refining queries to acquire valuable knowledge. This process continues until sufficient external information is gathered, at which point the results are presented to the user. To this end, we develop a method for autonomously synthesizing reasoning-based decision-making instructions in iterative retrieval and fine-tuned the latest open-source LLMs. The experimental results indicate that Auto-RAG is capable of autonomous iterative interaction with the retriever, effectively leveraging the remarkable reasoning and decision-making abilities of LLMs, which lead to outstanding performance across six benchmarks. Further analysis reveals that Auto-RAG can autonomously adjust the number of iterations based on the difficulty of the questions and the utility of the retrieved knowledge, without requiring any human intervention. Moreover, Auto-RAG expresses the iterative retrieval process in natural language, enhancing interpretability while providing users with a more intuitive experience\footnote{Code is available at \url{https://github.com/ictnlp/Auto-RAG}.

• The paper introduces an autonomous iterative retrieval framework that enhances LLM decision-making.
• It leverages multi-turn dialogue and reasoning-based query refinement to reduce noise and improve retrieval quality.
• Experimental evaluations show superior performance across benchmarks like Natural Questions and TriviaQA.

Overview of "Auto-RAG: Autonomous Retrieval-Augmented Generation for LLMs"

The paper presents Auto-RAG, an advanced model for Retrieval-Augmented Generation (RAG) designed to improve the decision-making capabilities of LLMs through autonomous iterative retrieval. This model addresses the shortcomings of existing RAG approaches, which often rely on manually constructed rules or few-shot prompting, thereby enhancing both efficiency and robustness in handling complex queries.

Problem Statement

RAG models are instrumental in addressing knowledge-intensive tasks by augmenting a generation process with external information retrieval. Traditional approaches to RAG, however, face challenges such as noise in retrieved content and limitations in retrieving comprehensive information for complex queries. These challenges lead to an increased inference overhead and limit the model's ability to leverage the full reasoning capabilities of LLMs.

Methodology

Auto-RAG introduces a systematic approach to iterative retrieval by leveraging LLMs' inherent decision-making and reasoning attributes in a fully autonomous manner.

1. Iterative Retrieval Framework: Auto-RAG employs a multi-turn dialogue with the retriever, enabling the system to engage in iterative planning and query refinement. The iterative process concludes when sufficient information has been gathered, allowing the LLM to produce a final response.
2. Reasoning and Decision-Making: The model autonomously synthesizes reasoning-based decision-making instructions that guide the iterative retrieval process. This involves the creation of a reasoning paradigm that determines the need for further information based on relevance and utility, thereby improving efficiency while minimizing unwarranted information processing.
3. Fine-Tuning Process: Auto-RAG's capabilities are enhanced by fine-tuning the latest open-source LLMs on synthesized reasoning tasks, which guide the model to better manage the retrieval process and optimize its interactions with the retriever.

Experimental Evaluation

The authors demonstrate the effectiveness of Auto-RAG by evaluating its performance across six benchmarks: open-domain and multi-hop QA datasets, including Natural Questions and TriviaQA. Noteworthy experimental results include:

• Superior performance compared to standard RAG and other iterative retrieval models, achieving high scores across multiple benchmarks.
• Capabilities for autonomous retrieval adaptation based on question difficulty and retriever performance.
• Improved interpretability and user experience by expressing the retrieval process in natural language.

Implications and Future Directions

Auto-RAG sets a new standard in retrieval-augmented generation by fully integrating the decision-making capabilities of LLMs with iterative retrieval processes, effectively enhancing both accuracy and efficiency. The results suggest possible improvements in various applications, from open-domain question answering to more specialized tasks requiring complex reasoning chains. Future research directions may explore further diversification of iterative methods and fine-tuning approaches, as well as application to other AI tasks beyond question answering. Moreover, the development of faster mechanisms for query refinement and noise reduction in retrieval could yield additional performance gains, especially as retrieval corpora grow and become more heterogeneous.

In conclusion, Auto-RAG represents an innovative step towards more autonomous and effective RAG systems, with broader implications for AI's ability to process and integrate vast amounts of information with minimal human oversight.