Nvidia Nemo Datadesigner Datadesigner

Data Designer Docs Maintenance

Unified skill for adding, updating, moving, and removing pages on the NeMo Data Designer Fern docs site.

Current URL: docs.nvidia.com/nemo/datadesigner (see instances in fern/docs.yml). Source of truth for everything user-facing is fern/.

Scope Rule

ALL doc prose edits happen under fern/. The remaining docs/ directory is only for notebook source, generated Colab notebooks, docs scripts, and downloadable recipe scripts. docs/notebook_source/*.py remains canonical for notebook code. Concept pages, recipes, plugins, and Dev Notes prose live under fern/versions/latest/pages/.

Versioning Model

main contains only the latest Fern authoring docs. Historical release snapshots live on the CI-managed docs-website branch and must not be edited on main.

fern/versions/
├── latest.yml          ← authoring nav file
└── latest/pages/       ← authoring MDX tree

docs.yml on main registers only slug: latest. When you edit docs, edit latest/pages/ and latest.yml.

Dev Notes are patched into the published latest docs by CI. Add new posts to latest/pages/devnotes/posts/ and latest.yml on main.

For future Fern-native releases, do not copy page trees by hand on main. The release workflow snapshots latest into a frozen version on docs-website.

Layout at a Glance

fern/
├── README.md                  ← maintainer cheat sheet
├── docs.yml                   ← title, theme, versions:, redirects, custom-domain
├── fern.config.json           ← organization + fern-api version pin
├── main.css                   ← bundled NVIDIA theme CSS
├── assets/                    ← logos, favicon, recipe assets, devnote post images (shared)
├── images/                    ← /images/* references from MDX
├── styles/                    ← per-component CSS (notebook-viewer, authors, metrics-table, …)
├── components/                ← React/JSX MDX components
│   ├── NotebookViewer.tsx     ← renders converted .ipynb cells with outputs
│   ├── Authors.tsx            ← devnote bylines (uses devnotes/authors-data.ts)
│   ├── MetricsTable.tsx       ← benchmark tables w/ best-value highlight
│   ├── TrajectoryViewer.tsx   ← multi-turn tool-call traces (research dev notes)
│   ├── BadgeLinks.tsx         ← header shields (license, github, etc.)
│   ├── Tag.tsx, CustomCard.tsx, CustomFooter.tsx
│   ├── notebooks/             ← gitignored per-tutorial *.json + *.ts output
│   └── devnotes/              ← .authors.yml, authors-data.ts, per-post trajectory data
├── scripts/
│   └── ipynb-to-fern-json.py  ← .ipynb → fern/components/notebooks/*.{json,ts}
└── versions/
    ├── latest.yml             ← authoring navigation tree
    └── latest/pages/          ← authoring MDX content

URL Routing Rules

Fern's URL is computed from the section/page titles in the active version YAML, not the file path:

File system                                           Published URL
────────────────────────────────────────────────────  ────────────────────────────────────────
latest/pages/concepts/columns.mdx                     /concepts/columns
latest/pages/concepts/tool_use_and_mcp.mdx            /concepts/tool-use-and-mcp/overview
latest/pages/recipes/code_generation/text_to_sql.mdx  /recipes/code-generation/text-to-sql
latest/pages/devnotes/posts/text-to-sql.mdx           /dev-notes/text-to-sql-for-nemotron-super

Rules:

• Section title → kebab-case slug: "Dev Notes" → dev-notes, "Code Generation" → code-generation, "Tool Use & MCP" → tool-use-and-mcp.
• Page title → kebab-case slug: page: Text-to-SQL for Nemotron Super → text-to-sql-for-nemotron-super (the filename text-to-sql.mdx is irrelevant for routing).
• Subdirectories in the file path are dropped — devnotes/posts/foo.mdx becomes /dev-notes/<page-title> (no /posts/).

When in doubt, recompute by walking the page's position in the active version YAML and slugifying each title.

Operations

Add a Page

1. Pick the file location: fern/versions/latest/pages/<subdir>/<filename>.mdx. Underscores in filenames are fine — they don't affect the URL.

2. Write minimal frontmatter:

   ---
   title: "<Page Title>"
   description: "One-line SEO description (or empty string)"
   ---
   
   # <Page Title>
   
   <content>
   

3. Add an entry under the appropriate section: in fern/versions/latest.yml:

   - page: <Page Title>
     path: ./latest/pages/<subdir>/<filename>.mdx
   

4. The URL becomes /<section-slug>/<page-title-slug>. Update any cross-references in other MDX accordingly.

5. If the page is a Dev Note, add it to the Dev Notes section in latest.yml.

Update a Page

Editing prose is straightforward — change the MDX, save. No mirror step (one canonical tree).

For title changes, also update the page: value in latest.yml and fix any incoming links pointing at the old slugified title.

Move / Rename a Page (with Redirect)

1. git mv the file.

2. Update path: in latest.yml.

3. Update incoming links: grep -rn "<old-slug>" fern/versions/latest/pages/.

4. Add a redirect to fern/docs.yml:

   redirects:
     - source: "/nemo/datadesigner/<old-section>/<old-slug>"
       destination: "/nemo/datadesigner/<new-section>/<new-slug>"
   

   Note redirects use the path-prefixed form because instances[0].url includes /nemo/datadesigner.

Remove a Page

1. grep -rn "<filename-or-slug>" to find inbound links.
2. git rm the MDX file.
3. Remove the - page: entry from latest.yml.
4. Fix or remove all inbound links.
5. Add a redirect in docs.yml if the URL was public.

Worked Example: Add a How-To Under Concepts

Request: "Add a how-to about deduplicating generated rows."

1. Create fern/versions/latest/pages/concepts/deduplication.mdx:

   ---
   title: "Deduplication"
   description: "Strategies for removing duplicates from generated datasets"
   ---
   
   # Deduplication
   
   <content>
   

2. Add to fern/versions/latest.yml under the Concepts section:

   - page: Deduplication
     path: ./latest/pages/concepts/deduplication.mdx
   

3. Published URL: /concepts/deduplication on latest.

4. cd fern && fern check && fern docs dev to verify.

Worked Example: Rename with Redirect

Request: "Rename /concepts/seed-datasets to /concepts/input-datasets."

1. Update page: Seed Datasets → page: Input Datasets in latest.yml.

2. git mv fern/versions/latest/pages/concepts/seed-datasets.mdx fern/versions/latest/pages/concepts/input-datasets.mdx and update path:.

3. Add to docs.yml:

   redirects:
     - source: "/nemo/datadesigner/concepts/seed-datasets"
       destination: "/nemo/datadesigner/concepts/input-datasets"
     - source: "/nemo/datadesigner/latest/concepts/seed-datasets"
       destination: "/nemo/datadesigner/latest/concepts/input-datasets"
   

4. grep -rn "/concepts/seed-datasets" fern/versions/latest/pages/ and rewrite hits.

Content Guidelines

DataDesigner uses Fern-native MDX components directly. Do not use GitHub > [!NOTE] or MkDocs !!! note syntax — neither renders.

Callouts

<Tip>
Recipes assume an OpenAI provider. Configure with `data-designer config list`.
</Tip>

Cards

Landing/index pages use <CardGroup cols={N}> + <Card title="..." icon="..." href="...">. Icons are Font Awesome names (no :material-...: shortcodes — those are MkDocs and render as text).

<CardGroup cols={3}>
  <Card title="Tutorials" icon="book-open" href="/tutorials/the-basics">
    Step-by-step notebooks
  </Card>
  <Card title="Recipes" icon="hat-chef" href="/recipes/recipe-cards">
    Ready-to-use examples
  </Card>
</CardGroup>

Header Badges (Landing Pages Only)

For shields-style badges (license, GitHub, model providers) at the top of an index page, use <BadgeLinks> from fern/components/BadgeLinks.tsx:

import { BadgeLinks } from "@/components/BadgeLinks";

<BadgeLinks
  badges={[
    { href: "https://github.com/...", src: "https://img.shields.io/badge/...", alt: "GitHub" },
  ]}
/>

This wraps a row of badge images in a flex container so they sit side-by-side rather than stacking.

Page Layout

The landing pages use frontmatter layout: overview for a wider design with TOC + sidebar. Other valid values: guide (default), reference (full-width, no TOC), page (no TOC, no nav), custom (blank canvas). See Fern docs.

Frontmatter

Match the PR #581 style — keep frontmatter minimal:

---
title: "<Page Title>"        # required
description: ""              # required (may be empty string)
layout: overview             # optional — only on landing pages
---

Do not add position: (we use explicit nav order in the version YAML), date:, or authors: to frontmatter — Fern's runtime treats authors: as a JSX scope variable and explodes when a component tries to reference it. For dev notes, see "Dev Notes" below.

Dev Notes (Blog Posts)

Dev notes live under fern/versions/latest/pages/devnotes/posts/. They use the dev-notes kit components: <Authors>, <MetricsTable>, <TrajectoryViewer>, <ExpandableCode>, <CustomCard> (sources in fern/components/, CSS in fern/styles/).

Add the page and card to latest.yml. Release snapshots are handled by CI on docs-website.

Authors Byline

Author registry: fern/components/devnotes/.authors.yml (source of truth) + fern/components/devnotes/authors-data.ts (typed copy that Authors.tsx imports). Edit both together.

---
title: "<Post Title>"
description: "<one-line summary>"
---

import { Authors } from "@/components/Authors";

# <Post Title>

<Authors ids={["dnathawani", "ymeyer"]} />

<post body>

Pass IDs as an explicit array literal, not <Authors ids={authors} /> — the bare authors identifier doesn't exist in Fern's MDX render scope.

To add a new author:

1. Append to fern/components/devnotes/.authors.yml:

   newauthor:
     name: New Author
     description: Researcher at NVIDIA
     avatar: https://avatars.githubusercontent.com/u/<id>?v=4
   

2. Add the same entry to fern/components/devnotes/authors-data.ts (no auto-sync).

Benchmark Tables

Use <MetricsTable> instead of plain markdown tables when you have numeric model comparisons. Auto-highlights the best value per column.

import { MetricsTable } from "@/components/MetricsTable";

<MetricsTable
  headers={["Model", "BIRD EX (%)", "Val Loss (↓)"]}
  rows={[
    ["Baseline", 26.77, 1.309],
    ["Ours", 41.80, 1.256],
  ]}
  lowerIsBetter={[2]}
/>

lowerIsBetter takes column indices; everything else is "higher is better".

Trajectory Viewer

For multi-turn agent traces (search/open/find/answer), put the trajectory data in a typed module under fern/components/devnotes/<post-slug>/<example>.ts:

import type { TrajectoryViewerProps } from "@/components/TrajectoryViewer";

const trajectory: Omit<TrajectoryViewerProps, "defaultOpen"> = {
  question: "...",
  referenceAnswer: "...",
  goldenPassageHint: "⭐ = golden passage",
  summary: "Example: 31 turns, 49 tool calls",
  turns: [
    { turnIndex: 1, calls: [{ fn: "search", arg: "..." }] },
    // ...
    { turnIndex: 31, calls: [{ fn: "answer", body: "..." }] },
  ],
};

export default trajectory;

Then in the post:

import { TrajectoryViewer } from "@/components/TrajectoryViewer";
import trajectory from "@/components/devnotes/<post-slug>/<example>";

<TrajectoryViewer {...trajectory} defaultOpen />

The component is collapsible when summary is set; uncollapsed when defaultOpen is true. Don't hand-code trajectory HTML/CSS in MDX — the original <div class="trajectory-viz"> + <style> approach was migrated away because MDX runs the CSS body through acorn and explodes on { color: ... }.

<ExpandableCode> — caveat

The kit ships ExpandableCode for collapsible code dumps, but it currently breaks Fern's SSR runtime (the onClick + setTimeout in the copy button trips the React error boundary, page renders "Something went wrong!"). Use plain ```python fences for now. Component file is kept in fern/components/ for future use if Fern's runtime support changes.

Notebook Tutorials

Notebook tutorials are JS-driven via <NotebookViewer> rendering converted .ipynb data. The whole pipeline:

docs/notebook_source/*.py            (jupytext format — canonical source, edit here)
        │ make convert-execute-notebooks      # jupytext --execute (needs NVIDIA_API_KEY)
        ▼
docs/notebooks/*.ipynb               (executed; outputs captured)
        │
        │ make generate-fern-notebooks         # per-file prefers executed docs/notebooks/
        │                                      # otherwise converts notebook_source/*.py directly
        ▼
fern/components/notebooks/*.{json,ts} (gitignored; generated before preview/publish)

docs/colab_notebooks/*.ipynb is a separate committed output for "Open in Colab" links. It is generated by make generate-colab-notebooks, but it is not a Fern docs build input.

The .ts is what the wrapper MDX imports. Fern's bundler doesn't follow .json imports cleanly.

Make Targets

Install notebook docs dependencies first with make install-dev-notebooks. Docs setup pins to DOCS_PYTHON_VERSION ?= 3.13 to match the published docs builds. Override via DOCS_PYTHON_VERSION=3.14 make ... (or any other supported version) when needed.

The convert-execute-notebooks step loops per file so one notebook missing an API key does not prevent later notebooks from running. Any failure is reported after the loop and the make target exits non-zero.

Wrapper Page

Each notebooks/<name>.mdx wrapper is tiny:

---
title: "The Basics"
description: "Declare columns, generate your first dataset."
---

import { NotebookViewer } from "@/components/NotebookViewer";
import notebook from "@/components/notebooks/1-the-basics";

<NotebookViewer
  notebook={notebook}
  colabUrl="https://colab.research.google.com/github/NVIDIA-NeMo/DataDesigner/blob/main/docs/colab_notebooks/1-the-basics.ipynb"
/>

The converter (fern/scripts/ipynb-to-fern-json.py) auto-strips the leading Colab badge cell — <NotebookViewer> renders its own banner from the colabUrl prop. Don't manually re-add it.

MDX Gotchas (the ones that bit during migration)

Validate

Run from fern/:

fern check          # YAML + frontmatter + MDX validation
fern docs dev       # localhost:3000 hot-reload preview

fern check must pass before commit. The local broken-link checker has known false positives — it computes URLs from file paths instead of from slugified nav titles, so cross-section absolute links sometimes flag incorrectly. Spot-check by clicking through the dev server.

Commit & Preview

git add fern/
git commit -s -m "docs: <add|update|remove> <page-title>"

DCO sign-off (-s) is required by CONTRIBUTING. Use docs: prefix (matches recent commit history). Subject line ≤ 50 chars (hard limit 72).

PRs touching Fern docs inputs get an automatic Fern preview URL posted as a comment. Fork PRs still run checks, but hosted preview publishing is skipped because it requires deployment secrets.

Cutting a New Version Train

Do not copy page trees by hand on main. The release workflow copies latest/pages/ to a frozen vX.Y.Z/pages/ tree on docs-website, updates the published versions list, checks the generated Fern docs, and publishes from docs-website.

Debugging

When NOT to Use This Skill

• Editing Python source under packages/ — that's a code change, not a docs change.
• Adding a notebook tutorial's code: edit docs/notebook_source/<name>.py, not the converted .ipynb or the wrapper MDX.
• Editing dev note prose: edit the MDX under fern/versions/latest/pages/devnotes/posts/<name>.mdx.

Source: NVIDIA-NeMo/DataDesigner — distributed by TomeVault.