LightRAG: Graph-Based Retrieval-Augmented Generation
HKUDS/LightRAGLightRAG is an open-source retrieval-augmented generation framework that adds a knowledge graph over your documents instead of relying only on flat vector similarity. It comes out of an EMNLP 2025 paper from HKUDS and ships as the `lightrag-hku` Python package.
Reach for LightRAG if your documents have entities and relationships that plain vector search keeps missing, and you want graph-aware retrieval without standing up a heavyweight platform. Skip it if a simple vector store already answers your questions well, or if you need a polished, long-supported product rather than a research-origin framework.
The problem it solves
Standard RAG chunks documents, embeds them, and retrieves by vector similarity. That works for lookup questions but struggles when an answer depends on how entities relate across a corpus — vector search returns similar-sounding passages, not the connected facts. You lose the structure that would let the model reason across documents.
What is it?
LightRAG is an open-source (MIT) retrieval-augmented generation framework, published as 'LightRAG: Simple and Fast Retrieval-Augmented Generation' at EMNLP 2025 by HKUDS. It builds a knowledge graph from your documents and uses it during retrieval, combining graph structure with text so the model can pull connected facts, not just similar chunks. It's distributed as the Python package `lightrag-hku`.
Why it's getting attention
Graph-based RAG has become a hot area since Microsoft's GraphRAG, and LightRAG positions itself as a simpler, faster take on the same idea. An accepted EMNLP 2025 paper plus a permissive MIT license and a ready-to-install package gives it academic credibility and practical reach at once, which the README's Trendshift badge reflects.
Key features
- ✓Builds a knowledge graph from your documents to capture entities and relationships
- ✓Combines graph structure with text retrieval instead of flat vector similarity alone
- ✓Backed by a peer-reviewed EMNLP 2025 paper from HKUDS
- ✓Permissive MIT license, friendlier for commercial use than copyleft frameworks
- ✓Installable as the `lightrag-hku` Python package (Python 3.10)
- ✓Aimed at faster, lighter graph RAG than heavier GraphRAG implementations
Best use cases
- •Question-answering over a corpus where entities and their relationships matter
- •Building a knowledge assistant that reasons across connected documents, not isolated chunks
- •Prototyping graph RAG without deploying a large end-to-end platform
- •Research and experiments comparing graph-based retrieval against plain vector RAG
How to install / try
LightRAG is a Python 3.10 package published on PyPI as `lightrag-hku` (install with `pip install lightrag-hku`). The README notes the project uses `uv` for package management (install uv via `curl -LsSf https://astral.sh/uv/install.sh | sh` on Unix/macOS, or the PowerShell command on Windows). See the repo for the current setup steps.
How to use
You index your documents so LightRAG builds a knowledge graph, then query it to retrieve graph-connected context for an LLM to answer with. It relies on an LLM and embedding model that you configure. Full API details and examples are in the repository and the linked paper.
Strengths
- ✓Captures entity relationships that flat vector search misses
- ✓Grounded in a peer-reviewed EMNLP 2025 paper, not just a marketing README
- ✓MIT license is easy to adopt commercially
- ✓Positioned as lighter and faster than heavier GraphRAG stacks
Limitations & risks
- △Building a knowledge graph adds indexing cost and complexity over a plain vector store
- △It's a research-origin framework, so APIs and behavior can shift between versions
- △You still supply and pay for the LLM and embedding models it calls
- △Graph RAG helps most on relationship-heavy corpora; for simple lookups it may be unnecessary overhead
Alternatives
Who should try it — and who should skip
Developers and researchers building RAG over corpora where relationships between entities carry the answer, and who want graph-aware retrieval without a heavy platform. If plain vector search already serves your use case, or you need a vendor-backed product, LightRAG's research-framework nature may be a poor fit.
Frequently asked questions
Regular RAG retrieves by vector similarity over text chunks. LightRAG additionally builds a knowledge graph from your documents, so retrieval can follow entity relationships rather than only matching similar text.
Yes. It comes from the paper 'LightRAG: Simple and Fast Retrieval-Augmented Generation', accepted at EMNLP 2025, by HKUDS.
It's on PyPI as `lightrag-hku` (`pip install lightrag-hku`), targeting Python 3.10. The README notes the project uses uv for package management.
MIT, which is permissive and generally straightforward to use in commercial projects.