Skip to main content

The First LAS App — nemo3super

A zero-config private RAG chatbot powered by NVIDIA Nemotron 3 Super 120B. Drop your docs into a folder. Chat with them in your browser. Your data stays local.

📥 Download the latest release​

What It Is​

nemo3super is the reference implementation for AIoOS's first pillar (Memory). It proves the LAS thesis in the smallest possible form:

  • One binary. One double-click. One private knowledge base.
  • Drop .pdf / .docx / .xlsx / .pptx / .csv / .md / .txt / .html / .py / .json into docs/ — the watcher embeds them in ~2 seconds.
  • Ask anything. Get answers with source citations.
  • Your files never leave your machine. Only embedding vectors round-trip to NVIDIA's free-tier API (or, with DGX Spark, stay 100% local).

What It Proves​

Claim from AIoOS thesisHow nemo3super demonstrates it
"LAS means AI you own"MIT license. Fork it. Ship your own.
"Data stays local"Raw docs never uploaded. Only query-time embeddings.
"Zero config"Double-click Start.bat. First-run wizard handles venv + API key.
"Production quality at day 1"10 public releases in 24 hours, fully tested.
"Built with AI-native dev"Every commit shipped via Claude Code CLI — zero lines hand-written.

Shipping Stats​

  • 🚀 10 releases in the first 24 hours (v0.1.0 → v0.2.1)
  • 💻 ~700 lines of Python (all generated by Claude Code)
  • 📦 ~20 KB zip download (dependencies install on first run)
  • 🌍 Cross-platform — Start.bat (Windows) / Start.command (macOS/Linux)
  • 🔐 Privacy-first — your docs are your docs

Architecture​

  User drops file into docs/
              │
              ▼
  ┌──────────────────────┐
  │  watchdog Observer   │  ← daemon thread inside app.py
  │  (auto-index in 2s)  │
  └──────────┬───────────┘
             │
             ▼
  ┌──────────────────────┐
  │  NVIDIA Embeddings   │  ← nv-embedqa-e5-v5
  │  (chunks → vectors)  │
  └──────────┬───────────┘
             │
             ▼
  ┌──────────────────────┐
  │  ChromaDB (local)    │  ← persists between runs
  └──────────┬───────────┘
             │
             ▼ (on query)
  ┌──────────────────────┐
  │  Nemotron 3 Super    │  ← 120B, via build.nvidia.com
  │  (w/ retrieved ctx)  │     (or local NIM on DGX Spark)
  └──────────┬───────────┘
             │
             ▼
  ┌──────────────────────┐
  │  Gradio Web UI       │  ← WeChat-style bubbles
  └──────────────────────┘

Try It in 3 Minutes​

  1. Install Python 3.10+ from python.org
  2. Get a free NVIDIA API key at build.nvidia.com
  3. Download Nemotron-RAG.zip · extract · double-click Start.bat

First-run setup takes ~3 minutes. After that it's instant.

Fork It​

nemo3super is MIT. If you're building something LAS-native, use it as a starting point:

git clone https://github.com/AIoOS-67/nemo3super
cd nemo3super

Swap the model. Swap the UI. Swap the vector DB. The architecture is yours.

What's Next​

More apps are coming — and we want yours. If you ship a LAS app, let us know: we'll feature it here.

Contact us →