llm-wiki

name	llm-wiki
description	Obsidian vault as compiled knowledge base. Use when user says ingest, compile, query, lint, evolve, or asks questions that should come from a knowledge base. Also triggers on 'init kb', 'new topic', 'health check', 'what's missing'.
context	fork
model	sonnet
allowed-tools	Bash, Read, Write, Edit, Glob, Grep, Agent, WebFetch

Read kb.yaml from the vault root before doing anything. Missing? Call MCP vault.init with the topic name, then proceed.

MCP Tools

All vault operations go through MCP. Filesystem fallback when Obsidian is closed -- same interface, direct file I/O.

Method	Does what
vault.init	Scaffold topic: raw/, wiki/, schema/, Log.md, kb.yaml
vault.read / vault.create / vault.modify / vault.append / vault.delete	CRUD (writes are dry-run gated)
vault.search / vault.searchByTag / vault.searchByFrontmatter	Find notes
vault.graph / vault.backlinks	Link graph
vault.batch	Multiple ops in one round-trip
vault.lint	Structural health check
vault.list / vault.stat / vault.exists	Navigation

kb_meta.py (deterministic, zero LLM cost)

KB_META = python /path/to/obsidian-llm-wiki/kb_meta.py
$KB_META <command> <vault_root> <topic> [args...]

Command	What
init	Create _meta.json
diff	Hash-diff raw/ against compiled state
update-hash <file>	Mark a raw file as compiled
update-index	Rebuild _index.md from wiki/
check-links	Validate [[wikilinks]] resolve
vitality	Access-weighted article scores
log-access <article>	Record a read event

Never use an LLM for these. They are exact.

Workflows

Ingest

Trigger: "ingest", "add source", user provides URL/file/path.

1. kb_meta.py init (no-op if exists)
2. Source type: URL -> WebFetch. PDF -> extract text. File -> copy.
3. Save to raw/{type}/ with frontmatter: source, ingested, source_type, authority
4. kb_meta.py diff -- report count
5. Ask: "Compile now?"

Compile

Trigger: "compile", "update wiki".

Gate: kb_meta.py diff -- if new and changed are both empty, stop.

Config: Read {topic}/schema/compile-config.yaml. Missing? Use templates/base/compile-config.yaml. Config has the extraction prompt, coverage thresholds, and parallelism settings.

Existing state: Read wiki/_index.md to get existing concept names (prevents duplicate creation).

Extract: Per changed source, dispatch a subagent (model per config model_tier):

• Send config's extraction.prompt + raw file content + existing concept names
• Subagent returns JSON: { summary, concepts[], relationships[] }
• Invalid JSON? Skip source, warn, continue.

Write summaries: One per source at wiki/summaries/{source-slug}.md. Overwrite if exists.

---
source: "[[raw/{source_file}]]"
compiled: {date}
---
# {title}
{one_liner}
## Key Points
- {point1}
- {point2}
## Related Concepts
- [[concepts/{concept-slug}]] -- {one_liner}

Write concepts: At wiki/concepts/{name-slug}.md.

• New concept: write full article (overview + properties + sources).
• Existing concept: append new properties and source. Do NOT rewrite overview.

Coverage tags: Count sources referencing each concept. Per config thresholds (default: 1=low, 2=medium, 4+=high).

Metadata: For each compiled source:

$KB_META update-hash "$VAULT" "$TOPIC" "raw/{file}"
$KB_META update-index "$VAULT" "$TOPIC"
$KB_META check-links "$VAULT" "$TOPIC"

Report: sources compiled, summaries written, new/updated concepts, broken links, coverage distribution.

See templates/examples/ for real output from the brillm topic.

Query

Trigger: question about knowledge base content.

1. Read wiki/_index.md, pick 3-8 relevant articles
2. Read articles. kb_meta.py log-access for each
3. Synthesize with [[wikilink]] citations. State coverage confidence.
4. If external data was needed: suggest filing to raw/ for compilation

Lint

Trigger: "lint", "health check".

Deterministic (free):

• check-links -- broken wikilinks
• diff -- uncompiled sources
• vitality -- stale/dead articles

LLM (sonnet scan, opus judgment):

• Redundant concepts to merge
• Contradicting claims across articles
• Referenced concepts with no article

Output: wiki/_lint-{date}.md

Evolve

Trigger: "evolve", "what's missing".

1. Read _index.md + high-vitality articles
2. vault.graph -- find orphans, weak clusters
3. Identify: missing articles, shallow hubs, research questions
4. Present ranked suggestions. User picks, agent executes.

Formatting

• [[wikilinks]] for all vault refs. Never markdown links for internal.
• YAML frontmatter: sources, coverage, compiled date.
• [[link|display text]] when display differs from target.

Mistakes That Will Waste Your Time

1. Recompiling unchanged sources -- kb_meta.py diff first, always
2. Skipping update-index + check-links after compile -- index goes stale
3. Using [text](path) instead of [[wikilinks]] -- Obsidian won't resolve
4. Running LLM for diffing or link checking -- kb_meta.py does it in milliseconds
5. Editing wiki articles by hand -- edit raw/, recompile. Wiki is LLM-maintained.
6. Forgetting log-access on query reads -- vitality scoring goes blind