LongRAG - A robust dual-perspective retrieval framework jointly developed by Zhipu, Tsinghua University, and the Chinese Academy of Sciences.

What is LongRAG?

LongRAG, developed by a research team from Tsinghua University, the Chinese Academy of Sciences, and Zhipu, is a dual-perspective robust retrieval enhancement generation (RAG) framework for long-text question answering (LCQA). Based on four components—a hybrid retrieval system, an LLM-enhanced information extractor, a CoT-guided filter, and an LLM-enhanced generator—it effectively addresses the challenges of global contextual understanding and factual detail recognition in long-text question answering. LongRAG outperforms baseline models such as long-context LLM, advanced RAG systems, and Vanilla RAG on multiple datasets, demonstrating superior performance and robustness. LongRAG provides an automated fine-tuning data construction pipeline, enhancing the system's "instruction-following" capability and domain adaptability.

LongRAG's main functions

• Dual-view information processing: To understand and answer questions about the context of long texts from two perspectives: global information and factual details.
• Hybrid SearcherQuickly retrieve relevant information snippets from large amounts of data.
• LLM Enhanced Information ExtractorThe retrieved fragments are mapped back to the original long text paragraphs to extract global background and structural information.
• CoT Guide FilterUse Chain of Thought (CoT) to guide the model to focus on information relevant to the problem and filter out irrelevant content.
• LLM Enhancement GeneratorThe final answer is generated by combining global information with key factual details.
• Automated fine-tuning data construction: Build high-quality fine-tuning datasets based on automated processes to improve model performance on specific tasks.

LongRAG's technical principles

• Search Enhancement Generation (RAG)Based on the RAG framework, external knowledge is retrieved to assist the language model in generating answers.
• Integration of global and detailed informationThe system not only focuses on local factual details, but also integrates global information from long texts to provide more comprehensive answers.
• Mapping strategyIt maps the retrieved fragments back to the original long text, restoring contextual information and providing a more accurate background structure.
• Chain Thinking (CoT)Using CoT as a global clue guides the model to gradually focus on problem-related knowledge, thereby increasing the evidence density.
• Filtering strategyBased on global clues of CoT, irrelevant information fragments are filtered out, while key factual details are retained.

LongRAG's project address

• GitHub repository:https://github.com/QingFei1/LongRAG
• arXiv technical paper:https://arxiv.org/pdf/2410.18050

Application scenarios of LongRAG

• Customer service and supportIn the field of customer service, understanding and answering lengthy customer inquiries or historical interaction records allows for more accurate responses and solutions.
• Medical consultationIn the healthcare industry, this involves handling large volumes of patient records and medical literature, and answering complex questions from doctors or patients about diseases, treatments, and medications.
• Legal ConsultationIt helps legal professionals provide in-depth analysis and advice on legal issues based on the analysis of a large number of legal documents and cases.
• Education and ResearchIn the field of education, it serves as a supplementary tool to help students and researchers gain a deeper understanding of lengthy academic articles and research reports, and to answer research-related questions.
• Enterprise Decision SupportAnalyze lengthy documents such as market research reports and corporate annual reports to provide data support and insights for business decisions.