NemoDuo

A hybrid multi-agent research system powered by NVIDIA Nemotron 3.

Nemotron 3 Super (120B) acts as the strategic Planner — reasoning deeply, decomposing queries into subtasks, and synthesizing final answers. Nemotron 3 Nano 4B runs locally via Ollama as the fast Executor — handling web search, document reading, and summarization.

This architecture directly mirrors the edge+cloud deployment pattern Nemotron 3 was designed for.

User Query
    │
    ▼
┌─────────────────────────────────┐
│   Orchestrator                  │
│   Routes · Coordinates · Logs   │
└────────────┬────────────────────┘
             │
    ┌────────┴────────┐
    ▼                 ▼
┌──────────────┐  ┌──────────────┐
│   PLANNER    │  │   EXECUTOR   │
│ Nemotron 3   │──│ Nemotron 3   │
│ Super (API)  │◀─│ Nano (Local) │
└──────────────┘  └──────────────┘
       │                │
       │    ┌───────────┤
       │    ▼           ▼
       │  Web Search  Doc Reader
       │  Summarizer  Tool Calls
       ▼
Final Answer + Citations + Metrics

---

✨ Features

• Reasoning budget control — dial Planner thinking depth from 256 to 8192 tokens
• Live agent stream — watch Super and Nano "talk" in real time in the UI
• Metrics dashboard — tokens used, reasoning tokens, latency, estimated API cost per run
• No LangChain — clean custom agent loop, easy to read and extend
• One-command setup via Docker Compose
• Free web search via DuckDuckGo (no API key needed), optional Serper

---

🚀 Quickstart

1. Clone & configure

git clone https://github.com/hadyy01/nemoDuo.git
cd nemoDuo
cp .env.example .env
# Edit .env — add your NVIDIA_API_KEY

2. Pull Nano locally via Ollama

# Install Ollama: https://ollama.com
ollama pull nemotron3-nano-4b

3. Run with Docker Compose

docker compose up --build

Open http://localhost:8501

Or run locally without Docker

pip install -r requirements.txt
streamlit run ui/app.py

---

⚙️ Configuration

Variable	Default	Description
NVIDIA_API_KEY	—	Your NVIDIA NIM API key
PLANNER_MODEL	nvidia/nemotron-3-super-120b-a12b	Super model via NIM
EXECUTOR_MODEL	nemotron3-nano-4b	Nano model via Ollama
REASONING_BUDGET	2048	Planner thinking tokens
MAX_SUBTASKS	5	Max tasks Nano executes per query
SERPER_API_KEY	—	Optional: use Serper instead of DuckDuckGo

---

🏗️ Project Structure

nemoDuo/
├── agents/
│   ├── planner.py          # Super via NVIDIA NIM
│   ├── executor.py         # Nano 4B via Ollama
│   └── orchestrator.py     # Coordinates both agents
├── tools/
│   ├── web_search.py       # DuckDuckGo / Serper
│   ├── doc_reader.py       # URL + PDF ingestion
│   └── summarizer.py       # Nano-powered summarizer
├── core/
│   ├── config.py           # Typed config from .env
│   ├── context_manager.py  # Shared 1M-token context window
│   └── metrics.py          # Token + cost + latency logger
├── ui/
│   └── app.py              # Streamlit dashboard
├── examples/
│   └── sample_queries.md
├── docker-compose.yml
├── Dockerfile
└── requirements.txt

---

🤖 How the Agents Collaborate

1. Planner (Super) receives the query and uses its reasoning budget to produce a structured JSON plan — a list of ordered subtasks with types: search, summarize, read_url, extract

2. Executor (Nano) runs locally and executes each subtask sequentially — searching the web, reading pages, summarizing content — returning structured results

3. Planner (Super) synthesizes all results into a comprehensive, cited final answer

4. Metrics are logged to SQLite and surfaced in the sidebar

---

📊 Reasoning Budget

The REASONING_BUDGET controls how many tokens Nemotron 3 Super spends thinking before responding. This is one of Nemotron 3's key features — inference-time reasoning budget control.

Budget	Use Case
256–512	Fast, low-cost queries
2048	Balanced (default)
4096–8192	Deep research, complex multi-step reasoning

---

Built with Nemotron 3 · NVIDIA NIM · Ollama · Streamlit