local-rag

100% local RAG for Obsidian, Zotero, and Claude Code — powered by LightRAG + Ollama

A complete, privacy-first RAG stack that runs entirely on your machine. No API keys. No cloud. No data leaves your computer.

Why local-rag?

Architecture

┌─────────────────────────────────────────────────────┐
│                  YOUR DOCUMENTS                      │
│   Obsidian vault (.md)  │  Zotero storage (.pdf)    │
└──────────────┬──────────────────────────────────────┘
               │  ingestion pipeline (Python)
               ▼
┌─────────────────────────────────────────────────────┐
│                 LIGHTRAG SERVER                      │
│  • Knowledge graph (entities + relations)            │
│  • Vector index (NanoVectorDB)                       │
│  • LLM: Ollama → qwen2.5:14b                        │
│  • Embeddings: Ollama → nomic-embed-text             │
│  • REST API at http://localhost:9621                 │
└──────────────┬──────────────────────────────────────┘
               │
       ┌───────┴───────┐
       ▼               ▼
┌─────────────┐  ┌──────────────┐
│ Claude Code │  │   Obsidian   │
│  (MCP tool) │  │Smart Connect.│
└─────────────┘  └──────────────┘

Quickstart

1. Install Ollama and pull models

# Download Ollama from https://ollama.com
ollama pull qwen2.5:14b        # ~8GB — the LLM
ollama pull nomic-embed-text   # ~300MB — embeddings

2. Install LightRAG

pip install "lightrag-hku[api]"

3. Clone and configure

git clone https://github.com/Ricardo-Kaminski/local-rag
cd local-rag
cp config.example.yaml config.yaml
# Edit config.yaml with your paths

4. Start the stack

Windows:

start-rag.bat

Linux/macOS:

./start-rag.sh

5. Run initial ingestion

conda activate local-rag
python ingest.py

6. Configure Claude Code MCP

Add to your Claude Code settings.json (~/.claude/settings.json):

"mcpServers": {
  "lightrag": {
    "command": "python",
    "args": ["C:/path/to/local-rag/mcp_server.py"]
  }
}

Now use query_rag, insert_document, and rag_health directly in Claude Code.

Install via pip

pip install local-rag-stack

After installation, use the local-rag CLI:

local-rag ingest          # index all documents once
local-rag watch           # continuous indexing daemon
local-rag start           # start LightRAG server + watcher
local-rag mcp             # start MCP server (for testing)

Configure Claude Code to use the MCP server (add to ~/.claude/settings.json):

"mcpServers": {
  "lightrag": {
    "command": "local-rag",
    "args": ["mcp"]
  }
}

Claude Code Plugin

Install directly from Claude Code:

/plugin install local-rag

MCP Tools (7 total)

Using Claude API as LLM (optional)

Use Claude instead of Ollama for RAG responses. In config.yaml:

llm:
  provider: "claude"
  model: "claude-opus-4-6"
  api_key: ""   # or set ANTHROPIC_API_KEY env var

Note: Ollama is still required for embeddings (nomic-embed-text). Only the response LLM can be replaced with Claude.

Guides by Use Case

• Obsidian users — query your personal knowledge base
• Zotero users — search your academic PDF library
• Generic documents — any folder of PDFs, Word files, or Markdown

Hardware Requirements

Tested on Windows 11 with RTX 3060 12GB. CPU inference works but is significantly slower.

Just want the MCP server?

If you already have a LightRAG instance running, install the standalone MCP package:

pip install lightrag-mcp

See lightrag-mcp for details.

Contributing

See CONTRIBUTING.md. Bug reports and feature requests welcome via GitHub Issues.

License

MIT — see LICENSE