marginalia 1.1.1

Markdown vault quality scanner for Obsidian, academics, and documentation teams

marginalia

Markdown vault quality scanner for Obsidian, academics, and documentation teams.

Zero dependencies. Pure Python. Works on any Markdown vault.

Install

Prerequisites

You need Python 3.9 or newer. Check if you have it:

python --version

If you don't have Python, install it from python.org/downloads — check "Add Python to PATH" during installation.

Install marginalia

pip install marginalia

That's it. Verify it works:

marginalia scan --help

Install from source (for contributors)

git clone https://github.com/hale-bopp-data/marginalia
cd marginalia
pip install -e .

Quick start

marginalia catalog
marginalia quickstart ~/my-vault/ --write
marginalia scan ~/my-vault/
marginalia scan ~/my-vault/ --json

If you're new to the tool, start with catalog to see the capability map, then run quickstart to generate an operator blueprint (operator-blueprint.json + .md) with the next recommended flow for your vault.

---

Commands

catalog — Operator capability map

marginalia catalog
marginalia catalog --json

Shows the tool catalog grouped by operator goal: baseline, catalog, normalization, materialization, guardrails, and measurement.

quickstart — Guided operator flow + blueprint materialization

marginalia quickstart ~/my-vault/
marginalia quickstart ~/my-vault/ --write
marginalia quickstart ~/my-vault/ --write --output out/ops

Reads the current vault state, identifies the next best slice, and suggests the operational flow. With --write, it materializes:

• operator-blueprint.json
• operator-blueprint.md

This is the fastest way to answer: "what should I run next on this vault?"

scan — Quality scan

marginalia scan ~/my-vault/
marginalia scan ~/my-vault/ --require title,tags,status
marginalia scan docs/ ../wiki/          # multi-vault
marginalia scan ~/my-vault/ --json

Checks for:

• Missing or incomplete YAML frontmatter (title, tags, status, …)
• Empty sections — heading-hierarchy-aware: ignores sections with sub-headings, code blocks, templates, and archive files
• Broken internal links — with "did you mean?" suggestions
• Broken [[wikilinks]]

Add --tag to auto-tag files with issues for Obsidian review (see scan --tag below).

link — TF-IDF link suggestions

marginalia link ~/my-vault/             # preview suggestions
marginalia link ~/my-vault/ --apply     # write ## See also sections
marginalia link docs/ ../wiki/ --min-score 0.3 --max-links 5

Finds semantically related notes using TF-IDF cosine similarity. Scores boost notes that share tags or directories. Writes ## See also sections with relative [[wikilinks]] (dry-run by default).

fix — Automated fixes (4 Giri pipeline)

marginalia fix ~/my-vault/              # dry-run
marginalia fix ~/my-vault/ --apply --no-what-if

Four passes: normalise frontmatter → fix broken links → clean empty sections → standardise headings.

fix-tags — Migrate flat tags to namespaced

marginalia fix-tags ~/my-vault/                     # dry-run
marginalia fix-tags ~/my-vault/ --apply
marginalia fix-tags ~/my-vault/ --taxonomy my.yml   # custom taxonomy

tags — Tag Dictionary & Inventory (L0)

# Fast: read existing frontmatter, detect synonyms by pattern
marginalia tags ~/my-vault/
marginalia tags ~/my-vault/ --out tag-dictionary.json

# Full: LLM reads each page, suggests tags with reasoning
marginalia tags ~/my-vault/ --analyze --out tag-inventory.json
marginalia tags ~/my-vault/ --analyze --taxonomy taxonomy.yml --out tag-inventory.json

Fast mode (default): reads existing frontmatter tags, counts usage, detects synonym candidates by name similarity.

Analyze mode (--analyze): for each page, the LLM reads the content and suggests tags with reasoning (why this tag?). The inventory records {file, existing_tags, suggested: [{tag, reason}]} for every page. Tags with similar reasons across different pages = synonyms.

Rationalize mode (--rationalize): the LLM sees the full tag landscape across all files and proposes merges — non-canonical domains → canonical, zombie namespaces → canonical, flat tags → namespaced. Returns proposed YAML merges ready to paste into the taxonomy.

Designed for a 3-level tag lifecycle:

1. tags --analyze (L0) — per-page inventory with reasoning ("what's there and why?")
2. tags --rationalize (L0→L1) — global rationalization across all tags
3. Edit taxonomy YAML (L1) — curate synonyms in merges: section
4. fix-tags --taxonomy (L2) — apply normalisation across the vault

scan --tag + untag — Obsidian review workflow

# Tag files with issues so you can find them in Obsidian
marginalia scan ~/my-vault/ --tag

# In Obsidian: search  tag:quality/review-needed

# After manual review, remove the tags
marginalia untag ~/my-vault/ --apply

Every scan with issues prints an Obsidian tip at the end showing how to find affected files.

check — Obsidian health check

marginalia check ~/my-vault/

Detects: .obsidian/ tracked in git, missing .gitignore, hierarchy too deep/flat, mixed-case dirs, unresolved [[wikilinks]], accidental Untitled.canvas.

discover — Hidden connections

marginalia discover ~/my-vault/ --json

Finds clusters of notes with overlapping tags that don't link to each other yet.

index — Generate MOC + indexes

marginalia index ~/my-vault/

Outputs: Map of Content (MOC), tag index, orphan index.

graph — Link graph JSON

marginalia graph ~/my-vault/ > graph.json

Returns: tag index, link graph, topology (hubs, authorities, orphans), tag clusters.

css — Tag colour snippets for Obsidian

marginalia css ~/my-vault/ > tags.css

ai — AI-powered analysis

marginalia ai review ~/my-vault/
marginalia ai tag ~/my-vault/
marginalia ai connect ~/my-vault/
marginalia ai frontmatter ~/my-vault/

Requires an API key for any OpenAI-compatible provider (see LLM Configuration below).

eval — Before/after RAG quality measurement

# Build a snapshot of current retrieval quality
marginalia eval snapshot ~/my-vault/ queries.yaml before.json

# After vault changes, take another snapshot
marginalia eval snapshot ~/my-vault/ queries.yaml after.json

# Compare
marginalia eval compare before.json after.json

Measures: top1_score, coverage, precision@K, recall@K. Verdict: IMPROVED / DEGRADED / NEUTRAL.

queries.yaml format:

queries:
  - text: "deploy to production"
    expected:
      - deploy.md
  - text: "semantic search qdrant"
    # no expected — coverage only

---

Config file

marginalia.yaml is auto-discovered in the current directory or vault root:

vaults:
  - docs/
  - ../wiki/
exclude:
  - node_modules/
  - .git/
  - archive/
min_score: 0.35
max_links: 5
top_k: 7
heading: "## See also"

---

Custom taxonomy

namespaces:
  course: [math, physics, history, literature, philosophy]
  type: [lecture-notes, essay, bibliography, summary, review]
  status: [draft, revision, final, published]
merges:
  notes: lecture-notes
  bib: bibliography

Then: marginalia fix-tags ~/vault/ --taxonomy my-taxonomy.yml

---

Obsidian plugin

An Obsidian plugin is available in obsidian-plugin/. It wraps the CLI with ribbon buttons and a results panel.

Build:

cd obsidian-plugin
npm ci
npm run build

Copy main.js, manifest.json, styles.css to your vault's .obsidian/plugins/marginalia/.

---

For whom?

• Students — Keep your thesis vault clean: frontmatter, links, tag structure
• Researchers — Map knowledge topology across hundreds of notes
• Documentation teams — Enforce quality gates on Markdown wikis
• Obsidian users — Find broken links, orphan notes, hierarchy issues, get automatic link suggestions

---

LLM Configuration

AI-powered commands (ai, tags --analyze, closeout --ai) require an API key. Set one of these environment variables:

Variable	Provider	Base URL (auto)
MARGINALIA_API_KEY	Any OpenAI-compatible	Set MARGINALIA_API_URL too
OPENROUTER_API_KEY	OpenRouter (default)	https://openrouter.ai/api/v1
DEEPSEEK_API_KEY	DeepSeek	https://api.deepseek.com
OPENAI_API_KEY	OpenAI	https://api.openai.com/v1

Optional:

• MARGINALIA_MODEL — model name (default: deepseek/deepseek-chat)
• MARGINALIA_API_URL — custom base URL (e.g., http://localhost:11434/v1 for Ollama)

All AI features are optional — marginalia works fully without any API key.

---

Testing

marginalia ships with 128 tests covering all core modules.

# Run all tests
python -m pytest tests/ -v

# Run a specific test module
python -m pytest tests/test_scanner.py -v
python -m pytest tests/test_linker.py -v
python -m pytest tests/test_fixer_giro7.py -v

Test coverage

Module	Tests	What's covered
test_scanner.py	Frontmatter parsing, tag extraction, broken link detection, empty sections, wikilinks, Giro 7 quality checks (summary_todo, stale_draft, empty_required_fields)
test_linker.py	TF-IDF vectorization, cosine similarity, relative link computation, tag overlap scoring
test_fixer_giro7.py	4-pass fixer pipeline, stale draft rules (path-based auto-resolution), frontmatter normalization
test_eval.py	RAG quality snapshots, before/after comparison, precision/recall metrics
test_closeout.py	Git data collection, session template generation
test_validators.py	YAML validation, taxonomy checks, retry logic

All tests are pure unit tests — no network, no filesystem side effects, no external services required.

---

Zero dependencies

marginalia uses only the Python standard library. No PyYAML, no external packages. Runs anywhere Python 3.9+ is installed.

Origin story

marginalia was born inside EasyWay, a data governance platform with 50+ AI agents and a growing wiki of 500+ Markdown files. As the wiki scaled, quality eroded: broken links, orphan pages, inconsistent tags, missing frontmatter. Manual reviews couldn't keep up.

We built marginalia to automate what humans forget: find every broken link, detect every missing tag, suggest every connection. Within days of deploying it, EasyWay's wiki went from 40% frontmatter coverage to 98%, broken links dropped to zero, and the tag taxonomy became consistent across all documentation.

The tool turned out to be useful far beyond our project — any Obsidian vault, research wiki, or documentation repo has the same problems. So we extracted it, removed all internal dependencies, and released it as a standalone product.

What it did for us, it can do for you.

Part of HALE-BOPP

Sovereign by design. Cloud by choice.

marginalia is part of the HALE-BOPP open-source ecosystem — portable, replicable tools for data and knowledge governance. Your vault quality runs where you decide, not where a vendor tells you.

• hale-bopp-db — Schema governance for PostgreSQL
• hale-bopp-etl — Config-driven data orchestration
• hale-bopp-argos — Policy gating and quality checks
• marginalia (this repo) — Markdown vault quality scanner

License

MIT