local-rag-stack 0.1.0

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

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?

	local-rag	Cloud RAG
Privacy	Your data stays on your machine	Sent to third-party servers
Cost	Free after hardware	Per-token API costs
Offline	Works without internet	Requires connectivity
Latency	Local inference	Network round-trips
Customization	Any model via Ollama	Limited to provider models

---

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)

Tool	Description
query_rag(question, mode?)	Query the knowledge base. Modes: local, global, hybrid (default)
insert_document(path)	Index a specific file by path
rag_health()	Check if LightRAG server is running
list_sources()	List configured sources with file counts
get_indexed_documents(limit?)	List documents already indexed
delete_document(doc_id)	Remove a document from the knowledge base
get_graph_labels()	List entity types in the knowledge graph

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

Component	Minimum	Recommended
RAM	8 GB	16 GB
GPU VRAM	0 (CPU only, slow)	8 GB (qwen2.5:7b) / 12 GB (qwen2.5:14b)
Disk	10 GB	20 GB

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