KAG

KAG (Knowledge Augmented Generation) is an open-source framework from the OpenSPG project for building logical reasoning and factual question-answering solutions over vertical-domain knowledge bases. It combines the OpenSPG knowledge graph engine with large language models. The latest release listed in the repository is v0.8.0, published in June 2025.

What It Is

KAG is a knowledge-enhanced LLM service framework for professional domains. The README says it aims to overcome the ambiguity of vector-similarity retrieval in traditional RAG and the noise that OpenIE introduces into GraphRAG. It supports logical reasoning and multi-hop fact Q&A over private knowledge bases that mix unstructured data, structured data, and expert rules.

How It Works

The framework has three parts: kg-builder, kg-solver, and kag-model. Only kg-builder and kg-solver are covered by the release described in the README.

• kg-builder builds an LLM-friendly knowledge representation based on the DIKW hierarchy. It uses layout analysis, knowledge extraction, property normalization, and semantic alignment to turn raw data and expert rules into a unified knowledge graph. It supports schema-free extraction and schema-constrained construction, with mutual indexing between graph structure and source text chunks.
• kg-solver is a logical-form-guided hybrid reasoning engine. Planning, reasoning, and retrieval operators turn natural-language questions into solving steps that can use exact-match retrieval, text retrieval, numerical calculation, or semantic reasoning.

Update: Version 0.8.0

The 0.8.0 release, dated 2025-06-27 in the README, adds private and public-network knowledge base modes, and public data sources such as LBS and web search can be integrated through MCP. Built-in index types include Outline, Summary, KnowledgeUnit, AtomicQuery, Chunk, and Table. Knowledge bases are now decoupled from applications, and KAG-powered inference Q&A can run inside agent workflows over MCP. The release also adapts the KAG-Thinker model. Version 0.7 added static and iterative task planning, Simple and Deep Reasoning modes, streaming output, and a Lightweight Build mode, which the README says cuts knowledge-construction token cost by 89%.

Setup Path

Users can run the product mode by launching the engine and dependent images with Docker Compose and opening the web UI at a local address. Developers can instead install the toolkit with pip from a cloned repository on Python 3.10 or later. The project is licensed under Apache 2.0.