lightrag

Graph-based retrieval augmented generation improves document Q&A by building an entity and relationship knowledge graph from source documents, enabling entity-aware context retrieval instead of relying solely on naive text similarity for answers.

You can extract entities and relationships for knowledge graph search by inserting multimodal documents like PDFs, images, and tables into LightRAG, which automatically builds a structured graph for entity-aware retrieval and Q&A.

Yes, you can store your RAG knowledge graph in PostgreSQL or Neo4J. LightRAG supports pluggable graph storage backends including Neo4J, MongoDB, PostgreSQL, and the default nano-vectordb for different production needs.

Graph-backed search offers naive, local, global, hybrid, and mix query workflows to balance retrieval speed and answer quality. Hybrid retrieval is configured by default when inserting and querying documents.

Yes, you need to configure an LLM for generation and matching vector embeddings for indexing and querying. You set one LLM provider API key in the environment file to start inserting documents and querying with graph-aware RAG.

Use global or hybrid query modes when you need broader entity-relationship context across documents. Naive retrieval suits simple text similarity matches, while hybrid and mix workflows combine local and global graph context to maximize answer quality.