Task-Classifying Retrieval Augmented Generation

Hallucination phenomena are inevitably present in large language models (LLMs) for generation tasks. Retrieval Augmented Generation (RAG) technology alleviates this issue to some extent by incorporating external retrieved documents. However, existing Retrieval-Augmented Generation (RAG) methods still face limitations in terms of task adaptability and the richness of retrieval inputs. To address these issues, this paper proposes an innovative RAG framework that enhances both task classification and high-quality document retrieval, aiming to improve the accuracy and robustness of generated results. Specifically, we design a task identification module capable of automatically categorizing user inputs into three types of tasks: information retrieval, text generation, and general question answering. This classification then guides the subsequent processing flow. At the retrieval stage, we introduce a HyDE-inspired hypothetical document generation strategy, which constructs task-relevant hypothetical answers to improve the alignment between retrieved documents and user intent. To improve the quality of generation, this article introduces a task adaptive reflection and generation mechanism, implementing differentiated reflection processes and generation strategies for different task types, thereby improving the accuracy and completeness of the model’s response in multi task scenarios. Experimental results demonstrate that the proposed approach achieves state-of-the-art performance across multiple natural language processing tasks, particularly excelling in complex question answering and open-domain information retrieval scenarios.