- Category
- AI Tools
- Rank
- No. 203Tools index
- Pricing
- Open Source
- Type
- TOOL
- Builder
- hkuds
- GitHub
- 39.5k stars
- Latest release
- v1.5.7
- Date
About
Simple, fast retrieval-augmented generation system from HKUDS that combines vector search and graph-based knowledge representation. EMNLP 2025.
What it does
LightRAG turns documents into a queryable knowledge base. Its ingestion pipeline parses files, chunks content, extracts graph knowledge, and stores retrieval data. Queries can combine retrieved context with a configured language model, while the server and WebUI support document insertion, search, visualization, citations, and returned contexts.
Why it's ranked here
The project offers unusually broad deployment and storage choices without abandoning a direct Python package. It supports local files, several databases, multiple model providers, reranking, document parsing, an API, and a WebUI. The verdict is positive for teams wanting a configurable RAG platform, tempered by meaningful setup, dependency, and security responsibilities.
What's good
Storage is modular across key-value, vector, graph, and document-status roles. Parsing covers common Office formats and PDF, with local and service-backed engines sharing one dispatch model. Operators get an interactive configuration wizard, offline preparation guidance, authentication options, security checks, citations, tracing, evaluation support, and reproducible dependency locking.
Tradeoffs
The server binds to all interfaces by default, and endpoints are public until authentication is configured. Ollama-compatible routes require extra whitelist configuration to protect. MinerU and Docling need external services on cache misses. The complete offline dependency group excludes heavyweight model stacks, while the broad provider, parser, and database matrix increases operational complexity.
How to use it well
Choose it for document-grounded applications that need a Python library, hosted API, or operator-facing WebUI with replaceable storage and model providers. Configure models first, add only the databases you need, enable reranking deliberately, and run the security audit before network exposure. Do not treat its offline package as a complete distribution for local heavyweight models or Docling.
Technical notes+
pyproject.toml defines the lightrag-hku Python package for Python 3.10+, setuptools packaging, API, storage, provider, evaluation, observability, offline, and test extras, plus six console scripts. Makefile drives uv and Bun setup, environment configuration, validation, security auditing, backups, and Apple container operations. docs/ParserDebugCLI.md documents single-file parser debugging with flat sidecar output and raw-cache reuse. docs/ThirdPartyParser.md specifies a shared parser contract, registry discovery through the lightrag.parsers entry point, suffix validation, queue groups, and per-document failure isolation. docs/Reproduce.md supplies evaluation prompts, datasets, comparison tables, and reproduction steps, but its comparisons rely on model-judged answer preferences rather than a conventional deterministic benchmark.
Observed
- License
- MIT
- Primary language
- Python, requiring Python 3.10 or newer
- Packaging
- Published as lightrag-hku with setuptools metadata and optional API, offline, storage, provider, evaluation, observability, and test dependency groups
- Install surface
- Supports uv, pip, editable source installation, Docker Compose, and prepared offline installation
- Interfaces
- Python library, FastAPI server, WebUI, parser debug CLI, and six packaged console commands
- Platform support
- Package metadata declares operating-system independence; documentation covers Windows, Linux, macOS, Docker, and Apple Silicon container deployment
- Testing structure
- Pytest is configured for a tests directory with offline, integration, database, API, and PostgreSQL smoke markers
Read from README.md, Makefile, setup.py, pyproject.toml, docs/Reproduce.md, docs/UV_LOCK_GUIDE.md, docs/ParserDebugCLI.md, docs/DockerDeployment.md, docs/InteractiveSetup.md, docs/ThirdPartyParser.md, docs/OfflineDeployment.md.
What it can do
Retrieve relevant information using vector search
Query text or natural language question → Relevant document chunks or passages
Build knowledge graph from documents
Text documents or corpus → Graph-based knowledge representation
Generate answers using retrieved context
User query and retrieved knowledge → Generated text response
Combine vector and graph search results
Search query → Hybrid retrieval results
Index documents for retrieval
Document collection or text corpus → Searchable vector and graph indexes
Extract entities and relationships from text
Raw text documents → Structured entities and relationship data
Tags
Tech Stack
Media
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.