hkuds/lightrag

LightRAG is a Python library for building Retrieval-Augmented Generation (RAG) systems that can answer questions about large document collections. RAG is a technique where an AI language model does not answer questions purely from its training data; instead, it first searches a document collection for relevant passages and then uses those passages as context to generate a grounded answer. LightRAG's distinguishing feature is that it structures its document knowledge as a knowledge graph rather than a flat list of text chunks. When you feed documents into LightRAG, it uses a language model to extract entities (people, places, concepts) and the relationships between them, building a graph where nodes are concepts and edges capture how they connect. When you ask a question, LightRAG can retrieve relevant information at two levels: local, focusing on specific entities and their direct neighbors in the graph, and global, reasoning about high-level patterns and themes across the entire document set. This dual-level retrieval means it can handle both narrow factual questions and broad, synthesizing questions better than approaches that only do flat text similarity search. Storage backends are pluggable: you can store the knowledge graph and vector embeddings in Neo4j, PostgreSQL, MongoDB, or OpenSearch, giving flexibility to choose the database that fits your infrastructure. There is also a web UI for inserting documents, querying, and visualizing the knowledge graph interactively. You would use LightRAG when building a question-answering system over a large corpus of documents, internal company knowledge bases, research literature, legal documents, or technical documentation, where you need the system to handle both specific detail questions and broad thematic questions well. It is a Python library, published as the package lightrag-hku, and presented as research at the EMNLP 2025 natural language processing conference.