Add Nemotron 3 Nano Omni Document Intelligence cookbook

[chiachihchen]

Summary

Add a self-contained, no-GPU Document Intelligence cookbook for Nemotron 3 Nano Omni that pairs it with Nemotron Parse on the hosted catalog endpoint (https://integrate.api.nvidia.com/v1):

• Layout extraction -- Parse turns each PDF page into a typed block tree (titles, sections, tables, picture bboxes).
• Picture transcription -- Nano Omni in Instruct mode (temperature=0.2, top_k=1) classifies + transcribes each cropped figure.
• Reading-order document -- text + transcriptions are stitched back into one Markdown bundle per PDF.
• Multi-page reasoning -- Nano Omni in Thinking mode (enable_thinking=True, temperature=0.6, top_p=0.95) answers free-form questions over the assembled documents.

Runs end-to-end on a free NVIDIA_API_KEY from build.nvidia.com -- no GPU, no Docker, no local model weights required.

Files

Under usage-cookbook/Nemotron-3-Nano-Omni/doc-intelligence-with-parse/:

• doc_intelligence_cookbook.ipynb -- 20-cell notebook with no-GPU intro callout, uv-aware install (with pip fallback), hard-fail key check, and chat_template_kwargs={"enable_thinking": True} for the QA call. Self-downloads the four demo PDFs from the public yubo2333/MMLongBench-Doc Hugging Face dataset on first run.
• README.md -- model overview, requirements, uv-based quick start, NIM Deploy-tab links, troubleshooting.
• .gitignore -- auto-downloaded PDFs, output artefacts, .env, .ipynb_checkpoints/.

Test plan

• All 8 code cells execute end-to-end against the catalog endpoint (~40 s, 0 errors).
• Figure parity: 4 PNG Parse layout overlays (cell 12), 5 PNG picture crops + 15 markdown Specialist transcripts (cell 14), 4 markdown QA results (cell 18), 1 inlined pipeline-flowchart PNG in markdown source.
• Both Deploy URLs return HTTP 200 on build.nvidia.com: Parse and Nano Omni.
• Cookbook subfolder is fully self-contained -- no external assets required at push time.