Retrieval Augmented Generation (RAG) systems are evolving from simple "retrieval + generation" concatenation towards complex knowledge systems integrating adaptive knowledge organization and multi-turn reasoning (such as DeepResearch and Search-o1). However, this increase in complexity leads to high engineering implementation costs for developers.
To address this, Tsinghua University THUNLP Lab, Northeastern University NEUIR Lab, OpenBMB, and AI9Stars jointly launched UltraRAG 2.0, enabling the rapid implementation of multi-stage reasoning systems with minimal code.
01 Development History of RAG 2.0
02 Application Scenarios of RAG 2.0
In the medical field, RAG 2.0 can quickly perform comprehensive analysis of patient medical records, imaging data, and self-reported information, providing personalized treatment suggestions, thereby alleviating the workload of doctors and improving diagnostic accuracy and patient satisfaction.
In the legal sector, by integrating laws and regulations, case precedents, and expert opinions, RAG 2.0 helps lawyers quickly find relevant legal basis. Furthermore, it can generate high-quality legal documents, reduce human errors, and improve the quality of legal services.
In the financial domain, it can quickly identify potential risk points, generate risk assessment reports, and provide corresponding risk management recommendations. This enhances both the bank's risk management level and customer trust.
03 Current Challenges of RAG 2.0
Limited Performance in Complex Reasoning Tasks
The accuracy of RAG generation results needs improvement in multi-hop reasoning or complex logical tasks. For example, when processing questions that require integrating information from multiple documents, insufficient information synthesis often occurs.
Retrieval Quality and Noise Issues
Reliance on vector databases can lead to low recall and hit rates, failing to retrieve relevant documents accurately. Outdated or inaccurate information within documents may mislead the model, causing deviations in the generated content and reducing credibility.
Security and Privacy Concerns
RAG requires access to external knowledge bases. Ensuring data security and privacy when sensitive data is involved is a significant challenge.
04 Recommended Frontier Papers on RAG 2.0
1. From Local to Global: A GraphRAG Approach to Query-Focused Summarization
【Key Insight】This paper proposes a graph-based RAG method for question answering tailored to private text corpora. It scales effectively to the breadth of user queries and the volume of indexed information by constructing an entity knowledge graph and generating community summaries to answer user questions.
【Methodology】The paper uses an LLM to construct a graph-based text index in two phases: first, deriving an entity knowledge graph from source documents, and second, pre-generating community summaries for all closely related entity groups. When a question is posed, each community summary is used to produce a partial response, and all partial responses are then aggregated into the final response delivered to the user.
【Experiment】In a class of global significance tasks on a dataset spanning 1 million tokens, the paper shows that the Graph RAG method provides significant improvements in the comprehensiveness and diversity of generated answers compared to naive RAG baselines.
2. Deep and Faithful Large Language Model Reasoning with Knowledge-guided Retrieval Augmented Generation
【Key Insight】This paper introduces Think-on-Graph 2.0, an improved Retrieval Augmented Generation framework. By aligning with and utilizing a knowledge graph as a navigation tool, it deepens and refines the RAG paradigm for information gathering and synthesis, enhancing the accuracy and reliability of large language model responses and demonstrating the potential of hybrid structured knowledge systems to significantly advance LLM reasoning.
【Methodology】The Think-on-Graph 2.0 framework uses knowledge graph-guided retrieval to align questions and leverage the graph for navigation, enhancing the depth and scope of information retrieval and ensuring logical consistency and semantic similarity.
【Experiment】The authors conducted extensive experiments on four public datasets, and the results showed that the Think-on-Graph 2.0 method outperforms baseline methods.