LightRAG — Fast Graph-Based RAG

LightRAG is a graph-based Retrieval-Augmented Generation system that extracts knowledge graphs from documents for entity-relationship-aware Q&A. Outperforms naive RAG, HyDE, and GraphRAG across multiple domains.

Use when: Building document Q&A systems, knowledge bases, semantic search with graph relationships, or multimodal RAG (PDFs, images, tables).

Source: github.com/HKUDS/LightRAG (EMNLP 2025, 13K+ stars)

Setup Workflow

1. Install Dependencies

cd tools/lightrag-plus
uv sync

Installs lightrag-hku and dependencies in isolated venv.

2. Configure LLM Provider

Add one of these to .env:

# OpenAI
OPENAI_API_KEY=sk-...

# Claude
ANTHROPIC_API_KEY=sk-ant-...

# Gemini
GEMINI_API_KEY=...

# Ollama (local)
OLLAMA_HOST=http://localhost:11434

3. Choose Storage Backend

Default (nano-vectordb) — no setup, good for prototyping:

rag = LightRAG(working_dir="./rag_storage")

Neo4J — advanced graph queries:

# Add to .env
NEO4J_URI=bolt://localhost:7687
NEO4J_USERNAME=neo4j
NEO4J_PASSWORD=...

MongoDB — unified storage:

# Add to .env
MONGODB_URI=mongodb://localhost:27017

PostgreSQL — SQL-based queries:

# Add to .env
POSTGRES_HOST=localhost
POSTGRES_PORT=5432
POSTGRES_USER=...
POSTGRES_PASSWORD=...
POSTGRES_DB=lightrag

Usage Patterns

Pattern 1: Core Library (Embedded)

Insert documents:

from lightrag import LightRAG, QueryParam
from lightrag.llm import openai_complete_if_cache, openai_embedding

rag = LightRAG(
    working_dir="./rag_storage",
    llm_model_func=openai_complete_if_cache,
    embedding_func=openai_embedding
)

# Insert
with open("document.txt") as f:
    rag.insert(f.read())

Query:

# Modes: naive, local, global, hybrid, mix
result = rag.query(
    "What is the main topic?",
    param=QueryParam(mode="hybrid")
)
print(result)

Pattern 2: Server Mode (Recommended for Production)

Start server:

cd tools/lightrag-plus
python -m lightrag.server --port 9621 --working-dir ./rag_storage

Query via REST API:

# Insert document
curl -X POST http://localhost:9621/insert \
  -H "Content-Type: application/json" \
  -d '{"text": "Your document text here"}'

# Query
curl -X POST http://localhost:9621/query \
  -H "Content-Type: application/json" \
  -d '{"query": "What is...?", "mode": "hybrid"}'

Web UI: http://localhost:9621 — knowledge graph visualization, node queries, subgraph filtering

Query Modes

Default: hybrid (no reranker) or mix (with reranker)

LLM/Embedding Configuration

LLM Requirements

• Minimum: 32B params, 32KB context
• Recommended: 64KB context
• Indexing: Non-reasoning models (faster)
• Query: Stronger models than indexing for best results

Embedding Model

• Recommended: BAAI/bge-m3, text-embedding-3-large
• Critical: Same model for indexing AND querying
• Dimension lock: PostgreSQL requires dimension at table creation — changing models = drop tables and recreate

Reranker (Optional)

• Boosts performance significantly when enabled with mode="mix"
• Recommended: BAAI/bge-reranker-v2-m3, Jina reranker API

LLM Provider Examples

OpenAI:

from lightrag.llm import openai_complete_if_cache, openai_embedding

rag = LightRAG(
    working_dir="./rag_storage",
    llm_model_func=openai_complete_if_cache,
    embedding_func=openai_embedding
)

Claude (Anthropic):

from lightrag.llm import anthropic_complete, openai_embedding  # use OpenAI for embeddings

rag = LightRAG(
    working_dir="./rag_storage",
    llm_model_func=anthropic_complete,
    embedding_func=openai_embedding
)

Ollama (Local):

from lightrag.llm import ollama_model_complete, ollama_embedding

rag = LightRAG(
    working_dir="./rag_storage",
    llm_model_func=ollama_model_complete,
    llm_model_name="llama3.2:latest",
    embedding_func=ollama_embedding,
    embedding_model_name="nomic-embed-text"
)

Gemini:

from lightrag.llm import gemini_complete, openai_embedding

rag = LightRAG(
    working_dir="./rag_storage",
    llm_model_func=gemini_complete,
    embedding_func=openai_embedding
)

Advanced Features

• Token tracking — rag.token_usage()
• Export KG data — rag.export_knowledge_graph()
• LLM cache — kv_store_llm_response_cache.json in working_dir
• Delete documents — rag.delete_document(doc_id) with auto KG regeneration
• Langfuse tracing — observability integration
• RAGAS evaluation — measure retrieval quality
• Multimodal RAG — integrate with RAG-Anything for PDFs, images, tables

Full docs: docs/AdvancedFeatures.md

File Structure

tools/lightrag-plus/
  pyproject.toml        ← Dependencies (lightrag-hku)
  README.md             ← Full documentation
  uv.lock               ← Dependency lock file
  .gitignore            ← Excludes .venv, rag_storage, build artifacts
  .venv/                ← Virtual environment (gitignored, recreated by `uv sync`)
  rag_storage/          ← Working directory (gitignored, auto-created on first use)

Commands

# Install
cd tools/lightrag-plus && uv sync

# Start server
cd tools/lightrag-plus && python -m lightrag.server --port 9621 --working-dir ./rag_storage

# Test insert/query via Python
cd tools/lightrag-plus && uv run python -c "
from lightrag import LightRAG, QueryParam
rag = LightRAG(working_dir='./rag_storage')
rag.insert('The capital of France is Paris.')
print(rag.query('What is the capital of France?', param=QueryParam(mode='naive')))
"

Platform Support

• Windows: Full support (requires Python 3.10+)
• macOS: Full support
• Linux: Full support

References

• GitHub: https://github.com/HKUDS/LightRAG
• Paper: https://arxiv.org/abs/2410.05779 (EMNLP 2025)
• PyPI: https://pypi.org/project/lightrag-hku/
• Core API: docs/ProgramingWithCore.md
• Server docs: docs/LightRAG-API-Server.md