KAG is a logical reasoning and Q&A framework based on the [OpenSPG](https://github.com/OpenSPG/openspg) engine and large language models, which is used to build logical reasoning and Q&A solutions for vertical domain knowledge bases. KAG can effectively overcome the ambiguity of traditional RAG vector similarity calculation and the noise problem of GraphRAG introduced by OpenIE. KAG supports logical reasoning and multi-hop fact Q&A, etc., and is significantly better than the current SOTA method.
The goal of KAG is to build a knowledge-enhanced LLM service framework in professional domains, supporting logical reasoning, factual Q&A, etc. KAG fully integrates the logical and factual characteristics of the KGs. Its core features include:
- Knowledge and Chunk Mutual Indexing structure to integrate more complete contextual text information
- Knowledge alignment using conceptual semantic reasoning to alleviate the noise problem caused by OpenIE
- Schema-constrained knowledge construction to support the representation and construction of domain expert knowledge
- Logical form-guided hybrid reasoning and retrieval to support logical reasoning and multi-hop reasoning Q&A
In the context of private knowledge bases, unstructured data, structured information, and business expert experience often coexist. KAG references the DIKW hierarchy to upgrade SPG to a version that is friendly to LLMs.
For unstructured data such as news, events, logs, and books, as well as structured data like transactions, statistics, and approvals, along with business experience and domain knowledge rules, KAG employs techniques such as layout analysis, knowledge extraction, property normalization, and semantic alignment to integrate raw business data and expert rules into a unified business knowledge graph.
This makes it compatible with schema-free information extraction and schema-constrained expertise construction on the same knowledge type (e. G., entity type, event type), and supports the cross-index representation between the graph structure and the original text block.
This mutual index representation is helpful to the construction of inverted index based on graph structure, and promotes the unified representation and reasoning of logical forms.
KAG proposes a logically formal guided hybrid solution and inference engine.
The engine includes three types of operators: planning, reasoning, and retrieval, which transform natural language problems into problem solving processes that combine language and notation.
In this process, each step can use different operators, such as exact match retrieval, text retrieval, numerical calculation or semantic reasoning, so as to realize the integration of four different problem solving processes: Retrieval, Knowledge Graph reasoning, language reasoning and numerical calculation.
* First, we refactored the KAG-Solver framework. Added support for two task planning modes, static and iterative, while implementing a more rigorous knowledge layering mechanism for the reasoning phase.
* Second, we optimized the product experience: introduced dual modes—"Simple Mode" and "Deep Reasoning"—during the reasoning phase, along with support for streaming inference output, automatic rendering of graph indexes, and linking generated content to original references.
* Added an open_benchmark directory to the top level of the KAG repository, comparing various RAG methods under the same base to achieve state-of-the-art (SOTA) results.
* Introduced a "Lightweight Build" mode, reducing knowledge construction token costs by 89%.
Please refer to [KAG usage (developer mode)](https://openspg.yuque.com/ndx6g9/cwh47i/rs7gr8g4s538b1n7#cikso) guide for detailed introduction of the toolkit. Then you can use the built-in components to reproduce the performance results of the built-in datasets, and apply those components to new busineness scenarios.
The KAG framework includes three parts: kg-builder, kg-solver, and kag-model. This release only involves the first two parts, kag-model will be gradually open source release in the future.
kg-builder implements a knowledge representation that is friendly to large-scale language models (LLM). Based on the hierarchical structure of DIKW (data, information, knowledge and wisdom), IT upgrades SPG knowledge representation ability, and is compatible with information extraction without schema constraints and professional knowledge construction with schema constraints on the same knowledge type (such as entity type and event type), it also supports the mutual index representation between the graph structure and the original text block, which supports the efficient retrieval of the reasoning question and answer stage.
kg-solver uses a logical symbol-guided hybrid solving and reasoning engine that includes three types of operators: planning, reasoning, and retrieval, to transform natural language problems into a problem-solving process that combines language and symbols. In this process, each step can use different operators, such as exact match retrieval, text retrieval, numerical calculation or semantic reasoning, so as to realize the integration of four different problem solving processes: Retrieval, Knowledge Graph reasoning, language reasoning and numerical calculation.
author={Liang, Lei and Sun, Mengshu and Gui, Zhengke and Zhu, Zhongshu and Jiang, Zhouyu and Zhong, Ling and Zhao, Peilong and Bo, Zhongpu and Yang, Jin and others},
title={KGFabric: A Scalable Knowledge Graph Warehouse for Enterprise Data Interconnection},
author={Yi, Peng and Liang, Lei and Da Zhang, Yong Chen and Zhu, Jinye and Liu, Xiangyu and Tang, Kun and Chen, Jialin and Lin, Hao and Qiu, Leijie and Zhou, Jun}
