Add agentic tool-calling use-case example for Nemotron 3 Super

[mvanhorn]

Summary

Adds a Jupyter notebook use-case example demonstrating Nemotron 3 Super's structured tool-calling capabilities for agentic workflows. Closes #105.

What's included:

• Jupyter notebook (agentic_tool_calling_tutorial.ipynb) with 7 progressive sections:
  1. Setup & configuration (NVIDIA API + vLLM options)
  2. Tool definition with OpenAI-compatible JSON schemas (5 simulated tools)
  3. Single tool call demonstration
  4. Multi-turn tool calling with conversation threading
  5. Autonomous agent loop with planning and multi-step execution
  6. Reasoning modes comparison (reasoning-off, regular, low-effort)
  7. Best practices (system prompts, boolean parameter handling per [BUG] Tool Calling: Boolean Parameters Returned as Python Strings ("True"/"False") Instead of JSON Format (true/false) #52)
• README with Models Used table, requirements, and quick start

Design decisions:

• All tools are simulated (no external API keys beyond NVIDIA) for reproducibility
• Uses NVIDIA API as primary endpoint with vLLM self-hosted as alternative
• Knowledge base focused on Nemotron architecture for contextually relevant demos
• Follows existing use-case-examples/ conventions (README + notebook pattern)

Research & Context

Project Context

• Stars: 510 | Language: Python + Jupyter | Active maintainers: 3
• Philosophy: Complete, reproducible pipelines with open data and training recipes

Related Issues & Community Signals

• [BUG] Tool Calling: Boolean Parameters Returned as Python Strings ("True"/"False") Instead of JSON Format (true/false) #52 - Tool calling boolean parameter bug (closed) - users actively using tool calling
• [Feature Request] Adaptive-K Routing: Dynamic Expert Selection Based on Router Entropy #58 - Feature request for Adaptive-K routing (open)
• PR Add Stock Analysis Agent use case example #15 - Stock Analysis Agent (stale since Nov 2025, 0 comments)
• Greptile evaluation found Super 3 returns useful code reviews in 12.5 seconds with 2 tool calls
• NemoClaw (NVIDIA's enterprise AI agent platform) launching at GTC March 16

Competitive Analysis

• DeepSeek: No official tool-calling examples
• Qwen: Has tool-calling examples in separate qwen-agent repository
• Llama: Meta's llama-recipes repo has function-calling examples
• Nemotron: Zero structured tool-calling examples despite being trained for it

What Gets Merged

• Cookbooks, use-case examples, docs improvements merge fast
• 18K+ lines merged in a single day for Super 3 launch
• Only 1 PR closed without merge (duplicate)
• Required: DCO sign-off, clear README, working notebook

This contribution was developed with AI assistance (Claude Code).