graph rag

Graph RAG combines knowledge graphs with retrieval to surface connected entities and relationships. **Standard RAG limitation**: Retrieves independent chunks, misses relationships across documents, can't answer "how does X relate to Y" well. **Graph RAG approach**: Build knowledge graph from documents (entities + relationships), for queries: identify relevant entities → traverse graph → retrieve connected information → generate answer with relationship context. **Construction**: Extract entities and relations using NER + relation extraction (LLM or specialized models), build graph database (Neo4j, NetworkX). **Query processing**: Parse query for entities → find in graph → expand neighborhood → retrieve relevant subgraph + associated text chunks. **Advantages**: Multi-hop reasoning (A→B→C connections), relationship-aware retrieval, entity disambiguation. **Microsoft's GraphRAG**: Hierarchical community summaries of entity clusters enable global queries. **Use cases**: Enterprise knowledge (people-projects-documents), research (papers-authors-topics), product catalogs (items-features-categories). **Complexity**: Graph construction expensive, maintenance overhead, query complexity. Powerful for relationship-heavy domains.