AutoSchemaKG (atlas-rag)
About this tool
A knowledge-graph construction framework from Associate Professor Yangqiu Song's group, associated with an ACL 2026 paper. Language models extract entity and event triples from documents and infer a schema through conceptualization, without a predefined ontology. The atlas-rag package supports graph construction, vector retrieval, multihop GraphRAG and Neo4j export. ATLAS billion-scale graphs are also published.
- Feature 1: Extract entity and event triples and infer schemas without predefined ontologies.
- Feature 2: atlas-rag supports OpenAI-compatible, Azure OpenAI, local Hugging Face and vLLM models.
- Feature 3: Vector retrieval and multihop GraphRAG, with Neo4j CSV export and hosted retrieval APIs.
- Feature 4: ATLAS-Wiki, ATLAS-Pes2o and ATLAS-CC together contain over 900 million nodes and 5.9 billion edges.
- Feature 5: PDF and Markdown processing, parallel graph construction and multiple languages.
Resource information
- Maintainer
- HKUST-KnowComp, Associate Professor Yangqiu Song's group.
- Type
- Tool / library
- Category
- Knowledge graphs & graph computing
- Runs on
- Tool or library
- Who it's for
- Researchers and knowledge-engineering teams building graphs and GraphRAG from papers, documents or web corpora, or experimenting with ATLAS graphs.
- License at time of listing
- MIT
- Source
- GitHub
