graphrag

You build a knowledge graph by ingesting documents, chunking them into TextUnits, and extracting entities, relationships, and claims. This graph-based private data reasoning approach connects dispersed information across documents into structured insights for multi-hop analysis.

Leiden clustering performs hierarchical community detection on extracted knowledge graph entities to generate multi-level summaries. This groups related entities and relationships into structured communities, enabling global reasoning and hierarchical summarization across large private datasets.

Yes, GraphRAG supports private datasets by extracting knowledge graphs from unstructured enterprise documents to enable reasoning over private data. It processes documents through ingestion, graph construction, and embedding generation without requiring external data exposure.

You query a knowledge graph by selecting global, local, or drift-style search modes to perform rich reasoning over extracted entities and hierarchical communities. Global queries summarize across communities, local queries focus on specific entities, and drift queries combine both approaches.

Vector search and graph embeddings combine in the GraphRAG pipeline by chunking data into TextUnits, generating embeddings for semantic retrieval, and storing them alongside the knowledge graph. This enables both similarity-based and relationship-based querying over private data.

The best way to extract entities and relationships is ingesting unstructured text, chunking it into TextUnits, and applying graph construction to identify entities, relationships, and claims. This builds a structured knowledge graph for reasoning over private datasets.