KAG: Boosting LLMs in Professional Domains via Knowledge Augmented Generation

TL;DR

Knowledge Augmented Generation (KAG) tackles the shortcomings of retrieval-augmented generation in professional domains by fusing knowledge graphs with vector retrieval through a comprehensive framework. It introduces an LLM-friendly representation (LLMFriSPG), mutual indexing between KG and text chunks, a logical-form guided hybrid solver, semantic knowledge alignment, and a unified KAG-Model with one-pass inference. Empirical results on multi-hop QA benchmarks show substantial end-to-end gains and industry deployments in E-Government and E-Health Q&A, demonstrating improved professionalism and reliability over traditional RAG. The work also discusses limitations related to computational cost and planning complexity and outlines future directions for efficiency, interpretability, and broader adoption in industry.

Abstract

The recently developed retrieval-augmented generation (RAG) technology has enabled the efficient construction of domain-specific applications. However, it also has limitations, including the gap between vector similarity and the relevance of knowledge reasoning, as well as insensitivity to knowledge logic, such as numerical values, temporal relations, expert rules, and others, which hinder the effectiveness of professional knowledge services. In this work, we introduce a professional domain knowledge service framework called Knowledge Augmented Generation (KAG). KAG is designed to address the aforementioned challenges with the motivation of making full use of the advantages of knowledge graph(KG) and vector retrieval, and to improve generation and reasoning performance by bidirectionally enhancing large language models (LLMs) and KGs through five key aspects: (1) LLM-friendly knowledge representation, (2) mutual-indexing between knowledge graphs and original chunks, (3) logical-form-guided hybrid reasoning engine, (4) knowledge alignment with semantic reasoning, and (5) model capability enhancement for KAG. We compared KAG with existing RAG methods in multihop question answering and found that it significantly outperforms state-of-theart methods, achieving a relative improvement of 19.6% on 2wiki and 33.5% on hotpotQA in terms of F1 score. We have successfully applied KAG to two professional knowledge Q&A tasks of Ant Group, including E-Government Q&A and E-Health Q&A, achieving significant improvement in professionalism compared to RAG methods.

Figures (10)

• Figure 1: The KAG Framework. The left side shows KAG-Builder, while the right side displays KAG-Solver. The gray area at the bottom of the image represents KAG-Model.
• Figure 2: LLMFriSPG:A knowledge representation framework that is friendly to LLMs. Instances and concepts are separated to achieve more effective alignment with LLMs through concepts. In this study, entity instances and event instances are collectively referred to as instances unless otherwise specified. SPG properties are divided into knowledge and information areas, also called static and dynamic area, which are compatible with decision-making expertise with strong schema constraints and document retrieval index knowledge with open information representation. The red dotted line represents the fusion and mining process from information to knowledge. The enhanced document chunk representation provides traceable and interpretable text context for LLMs.
• Figure 3: Hierarchical representation of knowledge and information.
• Figure 4: The Pipeline of KAG Builder for domain unstructured documents. From left to right, first, phrases and triples are obtained through information extraction, then disambiguation and fusion are completed through semantic alignment, and finally, the constructed KG is written into the storage.
• Figure 5: An Example of KAG-Builder pipeline