llm-wiki（LLM Wiki）

LLM Wiki

Agent-maintained personal knowledge base: ingest → query → lint → graph.

Ontology is opt-in. If wiki/ontology.yaml does not exist, all commands work exactly as described below. Ontology features are additive — see references/ontology-commands.md.

Project Layout

raw/          # Drop zone for unprocessed docs (originals stay after ingest)
wiki/
  index.md          # Catalog of all pages — update on every ingest
  log.md            # Append-only chronological record
  overview.md       # Living synthesis
  history.json      # Registry of all ingested sources
  ontology.yaml     # (OPTIONAL) project ontology schema
  synthesis-map.md  # Lightweight index of all saved query syntheses (append-only)
  originals/        # Read-only archive of source docs
  sources/          # One summary page per source (kebab-case.md)
  entities/         # People, companies, projects, products (TitleCase.md)
  concepts/         # Ideas, frameworks, methods, theories (TitleCase.md)
  syntheses/        # Saved query answers (kebab-case.md)
graph/              # graph.json + graph.html

Page templates and frontmatter schema → see references/templates.md. Wikilinks are case-sensitive: [[PageName]].

/wiki-ingest

Arguments (all optional):

• *(none)* → batch-ingest all files in raw/
• <file> → single file (e.g. raw/my-article.md)
• --from <folder> → use custom source folder instead of raw/
• --to <folder> → copy originals to PARA-structured subfolder (not wiki/originals/)
• --no-copy → skip copying originals; incompatible with --to
• --force-new → always create new entry (date-suffix slug if duplicate); mutually exclusive with --force-update
• --force-update → always overwrite existing entry; mutually exclusive with --force-new
• --no-interview → skip ontology Context Interview
• --batch-defaults → reuse first file's interview answers as defaults for the rest
• --summary-only → extract summary + metadata only; skip full-text conversion

For batch mode, --to PARA rules, file pre-processing (pdf/docx/pptx/xlsx), summary-only strategy, Context Interview → read references/ingest-advanced.md

Ingest Strategy (quick reference)

Duplicate Detection

Check wiki/history.json for the candidate slug before ingesting:

• --force-new: append today's date (e.g. my-article-20260421); proceed
• --force-update: overwrite existing entry; proceed
• Active slug, no flag: pause and ask — Update / New entry (<slug-YYYYMMDD>) / Cancel
• Deleted slug: treat as fresh ingest (reuse slug, set status: "active")

Agent-Based Ingest (full strategy)

1. Read the source file in full
2. Read wiki/index.md and wiki/overview.md. If wiki/ontology.yaml exists, run Context Interview (see references/ingest-advanced.md) unless --no-interview
3. Derive title and slug: a. Timestamp filename / empty title — if the filename matches YYYY-MM-DDTHH_MM_SS… or bare YYYY-MM-DD, or the source's title field is empty/whitespace:

- Read the first meaningful content lines (skip blank lines, lone URLs, timestamps) - Generate a concise descriptive title in the source's language (3–7 words) - Slug = kebab-case(title) + -YYYYMMDD date suffix (from file metadata or today) to guarantee uniqueness - If content is < 10 meaningful words or context is unclassifiable, use title 임시메모-YYYYMMDD and assign category 임시메모 b. Named filename — slug = kebab-case(filename, no extension); title = source's title field or first heading

1. Select template → see references/templates.md
2. Write wiki/sources/<slug>.md with source_file: wiki/originals/<slug>.md
3. Copy original to wiki/originals/<slug>.md (skip if --no-copy; apply PARA rules if --to)
4. Update wiki/index.md:

- Place the entry under the matching category ## section (create section if absent); maintain sections in this order: ## 기술/개발 · ## 업무/프로젝트 · ## 개인/생활 · ## 여행 · ## 임시메모 - Format: - [Title](sources/slug.md) — one-line summary (summary is mandatory — never omit or leave blank) - Classify by dominant content signal: - 기술/개발: code, tools, APIs, programming, software, hardware, dev workflows - 업무/프로젝트: work meetings, job-related memos, company projects, career - 개인/생활: accounts, IDs, passwords, addresses, family, finance, household - 여행: travel plans, trekking, transportation, accommodation, places - 임시메모: < 10 meaningful words, single snippet, or ambiguous context

1. Update wiki/overview.md — revise living synthesis if warranted
2. Create/update entity pages in wiki/entities/EntityName.md — when wiki/ontology.yaml exists, derive class: + properties: frontmatter from the matching class definition in the ontology; do not invent fields not declared there
3. Create/update concept pages in wiki/concepts/ConceptName.md — same ontology-derived frontmatter applies
4. Flag any contradictions with existing wiki content
5. Update wiki/history.json — always use Write tool (read existing → merge entry → overwrite). Schema → see references/ingest-advanced.md
6. Append to wiki/log.md: ## [YYYY-MM-DD] ingest | <Title>
7. Post-ingest: check broken [[wikilinks]], verify all new pages in index, print change summary. If wiki/ontology.yaml exists: print any schema warnings from this ingest as a ⚠ Schema warnings block (unknown class, missing required properties, unknown predicates) — these do not block ingest but should be resolved

/wiki-query

$ARGUMENTS = the question to answer

For structural filters (class:, type:, AND/OR/NOT, dotted field paths like context.phase) → read references/query-advanced.md

1. Synthesis-map lookup — if wiki/synthesis-map.md exists, read it in full; scan all entries and select the top 2–3 most semantically relevant past syntheses; read those wiki/syntheses/<slug>.md files in full
2. Read wiki/index.md to identify the most relevant pages
3. Read up to ~10 most relevant pages
4. If summaries lack sufficient detail, read the original from the page's source_file field
5. Present the answer with [[PageName]] wikilink citations; reference matched past syntheses with [[syntheses/slug]] links if they informed the answer
6. Include ## Sources listing every page and synthesis drawn from
7. End with: 💾 이 답변을 저장하려면: \/wiki-synthesize``

Empty wiki: *"The wiki is empty. Run /wiki-ingest <file> to add your first source."*

/wiki-synthesize

$ARGUMENTS (optional) = custom slug override (e.g. /wiki-synthesize my-custom-slug)

Saves the most recent /wiki-query answer from this session as a reusable synthesis.

1. Retrieve the query question and composed answer from the current session context
2. Generate a kebab-case slug from the query question (or use the provided override)
3. Extract entities, concepts, and relations from the answer
4. Write wiki/syntheses/<slug>.md (template: references/templates.md)
5. Append to wiki/synthesis-map.md (append-only — never edit past entries): ## YYYY-MM-DD | <slug> Query: <one-line question> Entities: Entity1, Entity2,... Concepts: Concept1, Concept2,... Relations: Subject→predicate→Object,... Preview: <one-sentence answer summary>
6. Append to wiki/log.md: ## [YYYY-MM-DD] query | <question summary>
7. Output: ✅ Saved synthesis: [[syntheses/<slug>]] — use /wiki-delete <slug> to remove if not needed

No recent query in session: *"No query found in this session. Run /wiki-query <question> first."*

/wiki-lint

python tools/lint.py
python tools/lint.py --save   # save to wiki/lint-report.md

Agent-based checks:

1. Orphan pages — pages with no inbound [[PageName]] links
2. Broken links — [[WikiLinks]] pointing to non-existent pages
3. Missing entity pages — names mentioned in 3+ pages but no wiki/entities/ page
4. Contradictions — read ## Contradictions sections; cross-check flagged claims
5. Stale summaries — source pages where last_updated < related concept/entity pages
6. Data gaps — questions the wiki can't answer; suggest source type to fill

Ontology checks (7–12), when wiki/ontology.yaml exists → read references/ontology-commands.md

After report, ask: "Save to wiki/lint-report.md?" Append to wiki/log.md: ## [YYYY-MM-DD] lint | Wiki health check

/wiki-graph

python tools/build_graph.py --open
python tools/build_graph.py --report --save

Agent-based fallback:

1. Glob every .md in wiki/. Build one node per page: {"id": "wiki/...", "label": "Title", "type": "source|entity|concept|synthesis", "color": "<hex>"} Colors: source #4CAF50 · entity #2196F3 · concept #FF9800 · synthesis #9C27B0
2. Grep [[wikilinks]]; resolve each to a node id and create an edge
3. Write graph/graph.json; inline its contents into graph/graph.html (template: references/graph-html.md)

Ontology-aware coloring, typed edges, and phase lanes → read references/ontology-commands.md

Append to wiki/log.md: ## [YYYY-MM-DD] graph | Knowledge graph rebuilt

/wiki-sources

1. Read wiki/history.json (if missing → *"No sources ingested yet. Run /wiki-ingest <file>."*)
2. Print table sorted by last_updated desc: Slug Title Ingested Last Updated Status
3. For each active source, list pages_created, entities_created, concepts_created

/wiki-update <slug>

1. Read wiki/history.json; error if slug absent or status: "deleted"
2. Find original at source_file; error if missing
3. Re-run Agent-Based Ingest steps 1–14 (update mode: do not re-archive original; overwrite source page; merge entity/concept pages)
4. Update wiki/history.json: set last_updated, refresh entity/concept lists
5. Append to wiki/log.md: ## [YYYY-MM-DD] wiki-update | <slug>
6. Output: ✅ Updated source: <slug>. Run /wiki-graph to regenerate the knowledge graph.

/wiki-delete <slug>

1. Read wiki/history.json; error if absent or already "deleted"
2. Show what will be deleted and ask for confirmation:

- wiki/sources/<slug>.md, source_file archive, orphaned entity/concept pages

1. After confirmation: delete source page, archive, and entity/concept pages not referenced by any other source. Update wiki/index.md and wiki/overview.md.
2. Set "status": "deleted" in wiki/history.json (preserve entry for audit trail)
3. Append to wiki/log.md: ## [YYYY-MM-DD] wiki-delete | <slug>
4. Output: ✅ Deleted source: <slug>. Run /wiki-lint to check for remaining broken links.

/wiki-ontology-* commands

Full flow for /wiki-ontology-init, /wiki-ontology-show, /wiki-ontology-validate → read references/ontology-commands.md

Key Gotchas

• wiki/originals/ is read-only — never modify archived files
• wiki/index.md must be updated every ingest — stale index breaks /wiki-query
• wiki/log.md is append-only — never edit past entries
• wiki/history.json: always use Write tool — read existing first, merge, then overwrite
• wiki/synthesis-map.md is append-only — never edit past entries; keep each entry under ~200 chars per line
• /wiki-delete never removes shared pages — cross-check references before deleting entity/concept pages
• Source slugs in sources: frontmatter must exactly match the source filename without extension
• All scripts run from the wiki project root (python tools/..., not from the scripts directory)