Rag Vault

RAG Vault

[License: MIT] [TypeScript] [MCP Registry]

Your documents. Your machine. Your control.

RAG Vault lets your AI coding assistant search your private documents, things like API specs, research papers, and internal docs. Everything runs locally and your data stays on your machine unless you choose to pull in content from a remote URL.

One command to run, minimal setup, privacy by default.

Why RAG Vault?

Security

RAG Vault comes with security built in:

• API Authentication: Optional API key via RAG_API_KEY
• Rate Limiting: You can throttle requests
• CORS Control: Restrict allowed origins
• Security Headers: Helmet.js protection

See SECURITY.md for complete documentation.

First-Time Setup Checklist

Before adding MCP config:

1. Install Node.js 20 or newer.
2. Pick a documents directory and set BASE_DIR to that path.
3. Make sure your AI tool process can read BASE_DIR.
4. Restart your AI tool after editing config.

Get Started Quickly

For Cursor

Add to ~/.cursor/mcp.json:

{  "mcpServers": {    "local-rag": {      "type": "stdio",      "command": "npx",      "args": ["-y", "github:RobThePCGuy/rag-vault"],      "env": {        "BASE_DIR": "/path/to/your/documents"      }    }  }}

Replace /path/to/your/documents with your real absolute path.

For Claude Code

Add to .mcp.json in your project directory:

{  "mcpServers": {    "local-rag": {      "type": "stdio",      "command": "npx",      "args": ["-y", "github:RobThePCGuy/rag-vault"],      "env": {        "BASE_DIR": "./documents",        "DB_PATH": "./documents/.rag-db",        "CACHE_DIR": "./.cache",        "RAG_EMBEDDING_DEVICE": "cpu",        "RAG_HYBRID_WEIGHT": "0.6",        "RAG_GROUPING": "related"      }    }  }}

Or add inline via CLI:

claude mcp add local-rag --scope user --env BASE_DIR=/path/to/your/documents -- npx -y github:RobThePCGuy/rag-vault

For Codex

Add to ~/.codex/config.toml:

[mcp_servers.local-rag]command = "npx"args = ["-y", "github:RobThePCGuy/rag-vault"]
[mcp_servers.local-rag.env]BASE_DIR = "/path/to/your/documents"

Install Skills (Optional)

If you want your AI to write better queries and make more sense of results, install the RAG Vault skills:

# Claude Code (project-level - recommended for team projects)npx github:RobThePCGuy/rag-vault skills install --claude-code
# Claude Code (user-level - available in all projects)npx github:RobThePCGuy/rag-vault skills install --claude-code --global
# Codex (user-level)npx github:RobThePCGuy/rag-vault skills install --codex
# Custom locationnpx github:RobThePCGuy/rag-vault skills install --path /your/custom/path

Skills teach Claude best practices for:

• Query formulation and expansion strategies
• Score interpretation. In boost mode, under 0.3 is a good match and over 0.5 is worth skipping. RRF mode scores by rank instead.
• When to use ingest_file vs ingest_data
• HTML ingestion and URL handling

Restart your AI tool, and start talking:

You: "Ingest api-spec.pdf"AI:  Successfully ingested api-spec.pdf (47 chunks)
You: "How does authentication work?"AI:  Based on section 3.2, authentication uses OAuth 2.0 with JWT tokens...

That's it. No Docker. No Python. No server infrastructure to manage.

Web Interface

RAG Vault has a web UI so you can manage your documents without touching the command line.

Launch the Web UI

npx github:RobThePCGuy/rag-vault web

Open http://localhost:3000 in your browser.

By default the web server binds to 127.0.0.1 (loopback), so it is reachable only from this machine. To use it from another device on your network, set RAG_BIND_HOST=0.0.0.0 — and set RAG_API_KEY as well, since the API is otherwise unauthenticated.

What You Can Do

• Upload documents: Drag and drop PDF, DOCX, Markdown, TXT, JSON, JSONL, and NDJSON files
• Search instantly: Type queries and see results with relevance scores
• Preview content: Click any result to see the full chunk in context
• Manage files: View all indexed documents and delete what you don't need
• Switch databases: Create and switch between multiple knowledge bases
• Monitor status: See document counts, memory usage, and search mode
• Export/Import settings: Back up and restore your vault configuration
• Theme preferences: Switch between light, dark, or system theme
• Folder browser: Navigate directories to select documents

REST API

The web server has a REST API you can hit directly. Set RAG_API_KEY to require authentication:

# With authentication (when RAG_API_KEY is set)curl -X POST "http://localhost:3000/api/v1/search" \  -H "Authorization: Bearer your-api-key" \  -H "Content-Type: application/json" \  -d '{"query": "authentication", "limit": 5}'
# Search documents (no auth needed if RAG_API_KEY isn't set)curl -X POST "http://localhost:3000/api/v1/search" \  -H "Content-Type: application/json" \  -d '{"query": "authentication", "limit": 5}'
# List all filescurl "http://localhost:3000/api/v1/files"
# Upload a documentcurl -X POST "http://localhost:3000/api/v1/files/upload" \  -F "[email protected]"
# Delete a filecurl -X DELETE "http://localhost:3000/api/v1/files" \  -H "Content-Type: application/json" \  -d '{"filePath": "/path/to/spec.pdf"}'
# Get system statuscurl "http://localhost:3000/api/v1/status"
# Health check (for load balancers)curl "http://localhost:3000/api/v1/health"

Reader API Endpoints

These endpoints let you read documents and find connections across them:

# Get all chunks for a document (ordered by index)curl "http://localhost:3000/api/v1/documents/chunks?filePath=/path/to/doc.pdf"
# Find related chunks for cross-document discoverycurl "http://localhost:3000/api/v1/chunks/related?filePath=/path/to/doc.pdf&chunkIndex=0&limit=5"
# Batch request for multiple chunks (efficient for UIs)curl -X POST "http://localhost:3000/api/v1/chunks/batch-related" \  -H "Content-Type: application/json" \  -d '{"chunks": [{"filePath": "/path/to/doc.pdf", "chunkIndex": 0}], "limit": 3}'

Remote Mode

RAG Vault can also run as an HTTP server so remote MCP clients like Claude.ai, Claude Desktop, or anything that supports Streamable HTTP or SSE can connect to it.

# Start remote server (default port 3001)npx github:RobThePCGuy/rag-vault --remote
# Custom portnpx github:RobThePCGuy/rag-vault --remote --port 8080

Stdio mode is unchanged. Just leave off --remote and everything works as before with Cursor, Claude Code, and Codex.

Connecting from Claude Desktop

Add to your Claude Desktop config:

{  "mcpServers": {    "rag-vault-remote": {      "type": "url",      "url": "http://localhost:3001/mcp"    }  }}

Or via Claude Code CLI:

claude mcp add --transport http rag-vault http://localhost:3001/mcp

Connecting from Claude.ai

For Claude.ai (Pro/Max/Team/Enterprise), add as a custom connector with URL https://your-host:3001/mcp. For local development, expose your server with a tunnel:

cloudflared tunnel --url http://localhost:3001

The remote server also binds to 127.0.0.1 (loopback) by default. To accept connections from other machines directly, set RAG_BIND_HOST=0.0.0.0. Set RAG_API_KEY for authentication whenever you expose the server beyond loopback (including via a tunnel). The server supports both Streamable HTTP (/mcp) and legacy SSE (/sse) transports, plus a health check at /health.

Real-World Examples

Search Your Codebase Documentation

You: "Ingest all the markdown files in /docs"AI:  Ingested 23 files (847 chunks total)
You: "What's the retry policy for failed API calls?"AI:  According to error-handling.md, failed requests retry 3 times     with exponential backoff: 1s, 2s, 4s...

Index Web Documentation

You: "Fetch https://docs.example.com/api and ingest the HTML"AI:  Ingested "docs.example.com/api" (156 chunks)
You: "What rate limits apply to the /users endpoint?"AI:  The API limits /users to 100 requests per minute per API key...

Build a Personal Knowledge Base

You: "Ingest my research papers folder"AI:  Ingested 12 PDFs (2,341 chunks)
You: "What do recent studies say about transformer attention mechanisms?"AI:  Based on attention-mechanisms-2024.pdf, the key finding is...

Search Exact Technical Terms

RAG Vault's hybrid search catches both meaning and exact matches:

You: "Search for ERR_CONNECTION_REFUSED"AI:  Found 3 results mentioning ERR_CONNECTION_REFUSED:     1. troubleshooting.md - "When you see ERR_CONNECTION_REFUSED..."     2. network-errors.pdf - "Common causes include..."

Pure semantic search would miss this. RAG Vault finds it.

How It Works

Document → Parse → Chunk by meaning → Embed locally → Store in LanceDB                         ↓Query → Embed → Vector search + BM25 → Fusion → Optional reranking → Results

Smart chunking: Splits by meaning, not character count. Keeps code blocks intact.

Hybrid search: Two fusion modes that combine vector similarity with BM25 keyword matching:

• Boost mode (default): BM25 boosts vector search distances multiplicatively. Simple and predictable.
• RRF mode (opt-in via RAG_SEARCH_MODE=rrf): Reciprocal Rank Fusion treats vector and BM25 as independent voters. This can surface documents that vector search alone would miss.

Cross-encoder reranking (opt-in): After the first pass, a cross-encoder model (Xenova/ms-marco-MiniLM-L-6-v2, ~23MB) scores each (query, passage) pair together for tighter relevance ranking. Turn it on with RAG_RERANKER_ENABLED=true.

Query expansion (opt-in): Generates reformulated queries to improve recall when searches are paraphrased or conceptual. Two backends: local template-based expansion (default, fully offline) or LLM-based HyDE through an external API. Turn it on with RAG_HYDE_ENABLED=true.

Quality filtering: Groups results by relevance gaps instead of arbitrary top-K cutoffs.

Local by default: Embeddings via Transformers.js. Storage via LanceDB. Network is only needed for initial model download or if you explicitly ingest remote URLs.

MCP tools included: query_documents, ingest_file, ingest_data, delete_file, list_files, status, feedback_pin, feedback_dismiss, and feedback_stats.

Supported Formats

Configuration

Environment Variables

Windows users: RAG_EMBEDDING_DEVICE=auto tries GPU providers (DirectML), which can fail if ONNX Runtime GPU binaries aren't available. If you see embedding initialization errors, set RAG_EMBEDDING_DEVICE=cpu in your MCP config for reliable operation. See the GPU acceleration FAQ for details.

One-command override (no .env edit):

# MCP modenpx github:RobThePCGuy/rag-vault --embedding-device cpu
# Web modenpx github:RobThePCGuy/rag-vault web --embedding-device dml
# Explicitly force auto detectionnpx github:RobThePCGuy/rag-vault --gpu-auto

Search Tuning

Cross-Encoder Reranking (opt-in)

Query Expansion / HyDE (opt-in)

Privacy note: The api backend sends query text to an external LLM endpoint, which breaks the "zero cloud" guarantee. The default rule-based backend is fully local.

Security (optional)

Advanced

Copy .env.example for a complete configuration template.

For code-heavy content, try:

"env": {  "RAG_HYBRID_WEIGHT": "0.8",  "RAG_GROUPING": "similar"}

Frequently Asked Questions

"env": {  "RAG_EMBEDDING_DEVICE": "cpu"}

"MODEL_NAME": "onnx-community/embeddinggemma-300m-ONNX"

Troubleshooting

Development

git clone https://github.com/RobThePCGuy/rag-vault.gitcd rag-vaultpnpm installpnpm -C web-ui install
# Install local git hooks (recommended, even for solo dev)pnpm hooks:install
# Fast local quality gate (backend + web-ui type/lint/format, deps, unused, build, unit tests)pnpm check:all
# Unit tests only (no model download required)pnpm test:unit
# Integration/E2E tests (requires model download/network)pnpm test:integration
# Buildpnpm build
# Run MCP server locally (stdio)pnpm dev
# Run MCP server locally (remote HTTP + SSE)pnpm dev:remote
# Run web server locallypnpm web:dev
# Release to npm (local, guarded)pnpm release          # patchpnpm release:minorpnpm release:majorpnpm release:dry

Test Tiers

• pnpm test:unit: deterministic tests for local/CI quality checks. Doesn't include model-download integration paths.
• pnpm test:integration: full integration and E2E workflows, including embedding model initialization.

Use RUN_EMBEDDING_INTEGRATION=1 to explicitly opt into network/model-dependent suites.

Release Strategy

• Releases are local and scripted via scripts/release-npm.sh.
• Supported bumps: patch, minor, major.
• The script runs dependency installs, pnpm check:all, and pnpm ui:build before touching version files.
• package.json and server.json versions only get updated after checks pass, and they're auto-restored if any later step fails.
• pnpm release:dry runs the full gate plus npm dry-run publish and always restores version files.

Project Structure

src/├── bin/             # CLI subcommands (skills install)├── chunker/         # Semantic text splitting├── embedder/        # Transformers.js wrapper├── errors/          # Error handling utilities├── explainability/  # Keyword-based result explanations├── flywheel/        # Feedback loop (pin/dismiss reranking)├── hyde/            # Query expansion + HyDE (LLM-based)├── parser/          # PDF, DOCX, HTML parsing├── query/           # Advanced query syntax parser├── reranker/        # Cross-encoder reranking (Transformers.js)├── server/          # MCP tool handlers + remote transport├── utils/           # Config, file helpers, process handlers├── vectordb/        # LanceDB + hybrid search (boost + RRF)└── web/             # Express server + REST API
web-ui/              # React frontend (Vite + Tailwind)

Documentation

• SECURITY.md: Security configuration and best practices
• .env.example: Complete environment variable template

License

MIT: free for personal and commercial use.

Acknowledgments

Built with Model Context Protocol, LanceDB, and Transformers.js.

Started as a fork of mcp-local-rag by Shinsuke Kagawa. Now it's its own thing. Huge credit to upstream contributors for the foundation, I've been iterating hard from there. Local-first dev tools, all the way.