AWS Labs Launches Apache 2.0 graphrag-toolkit for Multi-Hop GenAI Retrieval
Updated
Updated · KDnuggets · Nov 12
AWS Labs Launches Apache 2.0 graphrag-toolkit for Multi-Hop GenAI Retrieval
1 articles · Updated · KDnuggets · Nov 12
Summary
AWS Labs released graphrag-toolkit as an open-source Python package, showing how graph-enhanced RAG can answer multi-hop questions that plain vector search often misses.
The toolkit centers on lexical-graph, which builds a six-level knowledge graph—source, chunk, topic, statement, fact and entity—and uses traversal-based search to connect related facts before sending context to an LLM.
AWS said the package supports Neo4j, Amazon Neptune and other graph stores, plus Postgres with pgvector and other vector backends; the tutorial used local Neo4j and Postgres but still required Amazon Bedrock for model calls.
In a worked example using 3 Wikipedia articles, the build produced 47 chunks, 210 statements and 94 entities, illustrating how the system can link Commercial Crew, the ISS and Artemis across separate documents.
The launch positions GraphRAG as a fix for a structural RAG weakness: semantic similarity can find relevant text, but not the relationships needed for cross-document reasoning and more production-grade GenAI assistants.