KAG : A knowledge-augmented generation framework designed for building professional knowledge services.

KAG

KAG

Research Equipment Knowledge Management #Knowledge Augmentation #Generation Framework #Multi-hop Q&A #Knowledge Graphs #Specialized Fields Standard Picks Open Source

Overview :

KAG (Knowledge Augmented Generation) is a specialized framework for domain knowledge services, aimed at leveraging the strengths of knowledge graphs and vector retrieval to mutually enhance large language models and knowledge graphs. It addresses the significant gaps in vector similarity and knowledge reasoning relevance, as well as sensitivity to knowledge logic that are typical issues with Retrieval Augmentation Generation (RAG) techniques. KAG significantly outperforms methods like NaiveRAG and HippoRAG in multi-hop QA tasks; for instance, it achieves a 19.6% relative improvement in F1 scores on hotpotQA and a 33.5% increase on 2wiki. KAG has been successfully applied in two specialized knowledge Q&A tasks at Ant Group, including government and health Q&A, notably improving professionalism compared to RAG methods.

Target Users :

KAG targets developers and enterprise users, particularly teams that need to build knowledge services in specialized domains. It is suitable for users who require handling vast amounts of expertise, conducting complex reasoning, and implementing question-and-answer systems. KAG provides a complete framework and tools to help users convert domain knowledge into computable logical forms, thereby enhancing the professionalism and accuracy of Q&A systems.

Total Visits： 474.6M

Top Region： US(19.34%)

Website Views ： 72.0K

Use Cases

Government Q&A: Utilize the government knowledge graph built with KAG to accurately answer specialized questions regarding policies and regulations.

Health Q&A: Integrate medical knowledge through KAG to provide professional health consulting services.

Risk mining: In the finance sector, use expert rules defined by KAG to identify and analyze potential risks for applications and developers.

Features

LLM-friendly knowledge representation: Based on the DIKW hierarchy, it enhances SPG knowledge representation capability, compatible with both schema-less information extraction and schema-constrained professional knowledge construction.

Mutual indexing of knowledge graphs and original text snippets: Supports mutual indexing representation between graph structures and original text blocks, aiding the construction of inverted indexing based on graph structures, thus facilitating unified representation and reasoning of logical forms.

Logic-form guided hybrid reasoning engine: Includes operators for planning, reasoning, and retrieval, transforming natural language questions into problem-solving processes that combine language and symbols.

Application of domain knowledge scenarios: By defining expert rules and business data, KAG enables effective reasoning processes in scenarios such as risk mining.

Integration with OpenSPG conceptual modeling: Reduces the difficulty of converting natural language into graph queries, facilitating natural language Q&A in domain applications.

Scalability: KAG allows developers to extend the implementation of kag-builder and kag-solver to meet specific requirements.

Support for custom models and services: KAG supports integration with MaaS APIs compatible with OpenAI services as well as local models.

How to Use

1. Set up the environment and dependencies: Install software such as Docker and Docker Compose according to system version requirements.

2. Download and start the service: Use the curl command to download the docker-compose.yml file and start the service with Docker Compose.

3. Access the product: Enter the default URL http://127.0.0.1:8887 in your browser to view the product guide.

4. Install KAG: For developers, create a conda environment, clone the code, and install KAG following the guide.

5. Use the toolkit: Refer to the quick start guide to reproduce performance results using built-in components and apply them to new business scenarios.

6. Extend KAG capabilities: If the built-in components do not meet requirements, developers can extend KAG based on the KAG-Builder Extension and KAG-Solver Extension.

7. Adapt custom models: KAG supports integration with MaaS APIs compatible with OpenAI services and also facilitates integration with local models.