KAG

Why we included this project

Most question-answering setups over private or domain-specific data start with vector similarity retrieval, and that works until queries demand multi-hop reasoning or exact factual answers. KAG takes a different route: it pairs the OpenSPG knowledge graph engine with LLMs and converts natural-language questions into structured logical-form steps instead of a single similarity search. The framework folds unstructured text, structured records, and expert business rules into one schema-constrained knowledge graph, then runs a mixed reasoning engine that can pick between text retrieval, graph reasoning, numeric computation, and language reasoning at each step. That suits vertical professional domains where answers need to be grounded in source references and a plain RAG pipeline is prone to confident but wrong output. Teams with knowledge-graph or data-modeling experience will get the most value; those after a zero-setup plug-in RAG layer will find it heavier than they want.