wiki-export

name

wiki-export

description

Export the Obsidian wiki's knowledge graph to structured formats for use in external tools. Use this skill when the user says "export wiki", "export graph", "export to JSON", "export to Gephi", "export to Neo4j", "export to Postgres", "export to SQL", "graphml", "visualize wiki", "knowledge graph export", "export to OKF", "OKF bundle", "open knowledge format", "export as markdown bundle", or wants to use their wiki data in another tool. Outputs graph.json, graph.graphml, cypher.txt (Neo4j), postgres.sql (Postgres), and graph.html (interactive browser visualization) into a wiki-export/ directory at the vault root, plus an optional OKF (Open Knowledge Format) markdown bundle under wiki-export/okf/.

Wiki Export — Knowledge Graph Export

You are exporting the wiki's wikilink graph to structured formats so it can be used in external tools (Gephi, Neo4j, custom scripts, browser visualization).

Before You Start

1. Resolve config — follow the Config Resolution Protocol in llm-wiki/SKILL.md (inline @name override → walk up CWD for .env → global config → prompt setup). This gives OBSIDIAN_VAULT_PATH
2. Confirm the vault has pages to export — if fewer than 5 pages exist, warn the user and stop

Project Filter (optional)

If the user's invocation includes a project name — e.g. /wiki-export prismor, "export the prismor project", "export project:security" — activate project filter mode:

1. Extract the project name from the argument or phrase. Normalise: lowercase, strip the word "project".
2. Keep only pages where either condition holds:
   • The page id starts with projects/<name>/ (path-based match)
   • The page's tags array contains <name> (tag-based match)
3. Drop any edge where either endpoint was excluded.
4. Note the filter in the summary: (filtered: project:<name> — X of Y pages)
5. Set graph.graph.filter = "project:<name>" in the JSON output.

If both a project filter and a visibility filter are active, apply both (project filter first, then visibility filter on the remaining set).

Visibility Filter (optional)

By default, all pages are exported regardless of visibility tags. This preserves existing behavior.

If the user requests a filtered export — phrases like "public export", "user-facing export", "exclude internal", "no internal pages" — activate visibility filtered mode:

• Build a blocked tag set: {visibility/internal, visibility/pii}
• Skip any page whose frontmatter tags contain a blocked tag when building the node list
• Skip any edge where either endpoint was excluded
• Note the filter in the summary: (filtered: visibility/internal, visibility/pii excluded)

Pages with no visibility/ tag, or tagged visibility/public, are always included.

Step 1: Build the Node and Edge Lists

Glob all .md files in the vault (excluding _archives/, _raw/, _readouts/, .obsidian/, index.md, log.md, _insights.md). Apply any active filters (project and/or visibility) after collecting the full file list.

For each page, extract from frontmatter:

• id — relative path from vault root, without .md extension (e.g. concepts/transformers)
• label — title field from frontmatter, or filename if missing
• category — directory prefix (concepts, entities, skills, references, synthesis, projects, or journal)
• tags — array from frontmatter tags field
• summary — frontmatter summary field if present

This is your node list.

For each page, Grep the body for \[\[.*?\]\] to extract all wikilinks:

• Parse each [[target]] or [[target|display]] — use the target part only
• Resolve the target to a node id (normalize: lowercase, spaces→hyphens, strip .md)
• Skip links that point outside the node list (broken links)
• Each resolved link becomes an edge: {source: page_id, target: linked_id, relation: "wikilink", confidence: "EXTRACTED"}
• If the linking sentence ends with ^[inferred] or ^[ambiguous], override confidence accordingly

Typed edge enrichment: After building the wikilink edge list, read each page's relationships: frontmatter block. For each {target, type} entry:

• The target YAML value is a quoted wikilink string such as "[[concepts/lstm]]". Strip the surrounding [[ and ]] characters, then apply the same normalization (lowercase, spaces→hyphens, strip .md) to get the node id.
• Skip entries whose resolved target is not in the node list (broken link)
• If an edge for this (source, target) pair already exists, override its relation field with the typed value (e.g., "contradicts") and set typed: true
• If no edge exists yet for this pair, add one: {source: page_id, target: target_id, relation: <type>, confidence: "EXTRACTED", typed: true}

This means relation: "wikilink" is the default for plain untyped links; a relationships: entry promotes it to a named semantic type. Edges that originated from both a body wikilink and a relationships: entry keep a single record — the typed version wins.

This is your edge list.

Step 2: Assign Community IDs

Group pages into communities by tag clustering:

• Pages sharing the same dominant tag belong to the same community
• Dominant tag = the first tag in the page's frontmatter tags array
• Pages with no tags get community id null
• Number communities starting from 0, ordered by size descending (largest community = 0)

This enables community-based coloring in the HTML visualization and tools like Gephi.

Step 3: Write the Output Files

Create wiki-export/ at the vault root if it doesn't exist. Write all five files:

---

3a. graph.json

NetworkX node_link format — standard for graph tools and scripts:

{
  "directed": false,
  "multigraph": false,
  "graph": {
    "exported_at": "<ISO timestamp>",
    "vault": "<OBSIDIAN_VAULT_PATH>",
    "total_nodes": N,
    "total_edges": M
  },
  "nodes": [
    {
      "id": "concepts/transformers",
      "label": "Transformer Architecture",
      "category": "concepts",
      "tags": ["ml", "architecture"],
      "summary": "The attention-based architecture introduced in Attention Is All You Need.",
       
    
  
   
    
       
       
       
       
    
    
       
       
       
       
       
    
  

---

3b. graph.graphml

GraphML XML format — loadable in Gephi, yEd, and Cytoscape:

<?xml version="1.0" encoding="UTF-8"?>
<graphml xmlns="http://graphml.graphdrawing.org/graphml">
  <key id="label" for="node" attr.name="label" attr.type="string"/>
  <key id="category" for="node" attr.name="category" attr.type="string"/>
  <key id="tags" for="node" attr.name="tags" attr.type="string"/>
  <key id="community" for="node" attr.name="community" attr.type="int"/>
  <key id="relation" for="edge" attr.name="relation" attr.type="string"/>
  <key id="type" for= = =/>
  
  
    
      Transformer Architecture
      concepts
      ml, architecture
      0
    
    
    
      wikilink
      EXTRACTED
    
    
    
      contradicts
      contradicts
      EXTRACTED
    
  

Write one <node> per page and one <edge> per link. For typed edges (those where typed: true in the edge list), emit both <data key="relation"> with the semantic type value and <data key="type"> with the same value — this keeps relation readable for tools that already consume it while letting type-aware tools filter on the dedicated type key. Untyped wikilinks omit the <data key="type"> element entirely.

---

3c. cypher.txt

Neo4j Cypher MERGE statements — paste into Neo4j Browser or run with cypher-shell:

// Wiki knowledge graph export — <TIMESTAMP>
// Load with: cypher-shell -u neo4j -p password < cypher.txt

// Nodes
MERGE (n:Page {id: "concepts/transformers"}) SET n.label = "Transformer Architecture", n.category = "concepts", n.tags = ["ml","architecture"], n.community = 0;
MERGE (n:Page {id: "entities/vaswani"}) SET n.label = "Ashish Vaswani", n.category = "entities", n.tags = ["person","ml"], n.community = 0;
MERGE (n:Page {id: "concepts/lstm"}) SET n.label = "LSTM", n.category = "concepts", n.tags = ["ml","rnn"], n.community = 0;

// Relationships
// Untyped wikilinks use [:WIKILINK]
MATCH (a:Page {id: "concepts/transformers"}), (b:Page {id: "entities/vaswani"}) MERGE (a)-[:WIKILINK {relation: "wikilink", confidence: "EXTRACTED"}]->(b);
// Typed edges use the relationship type as the label (UPPERCASE)
MATCH (a:Page {id: "concepts/transformers"}), (b:Page {id: "concepts/lstm"}) MERGE (a)-[:CONTRADICTS {relation: "contradicts", confidence: "EXTRACTED"}]->(b);

Write one MERGE node statement per page, then one MATCH/MERGE relationship statement per edge. For typed edges, use the type value uppercased as the Cypher relationship label (e.g., contradicts → [:CONTRADICTS], derived_from → [:DERIVED_FROM]). Untyped wikilinks always use [:WIKILINK].

---

3d. postgres.sql

Plain SQL — loadable into any Postgres database (local, Supabase, RDS, Neon, …) with psql -f postgres.sql or a migration runner. Two tables: wiki_pages (nodes) and wiki_edges (links), with ON CONFLICT upserts so re-running the export is safe and idempotent, mirroring the MERGE semantics of cypher.txt.

-- Wiki knowledge graph export — <TIMESTAMP>
-- Load with: psql -d yourdb -f postgres.sql

CREATE TABLE IF NOT EXISTS wiki_pages (
  id        TEXT PRIMARY KEY,
  label     TEXT NOT NULL,
  category  TEXT,
  tags      JSONB NOT NULL DEFAULT '[]'::jsonb,
  summary   TEXT,
  community INT
);

CREATE TABLE IF NOT EXISTS wiki_edges (
  source     TEXT NOT NULL REFERENCES wiki_pages(id) ON DELETE CASCADE,
  target     TEXT NOT NULL REFERENCES wiki_pages(id) ON DELETE CASCADE,
  relation   TEXT NOT NULL DEFAULT 'wikilink',
  confidence TEXT,
  typed      BOOLEAN NOT NULL DEFAULT false,
  PRIMARY KEY (source, target, relation)
);

CREATE INDEX IF NOT EXISTS wiki_edges_source_idx ON wiki_edges(source);
CREATE INDEX IF NOT EXISTS wiki_edges_target_idx ON wiki_edges(target);

-- Nodes
INSERT INTO wiki_pages (id, label, category, tags, summary, community)
VALUES ('concepts/transformers', 'Transformer Architecture', 'concepts', '["ml","architecture"]'::jsonb, 'The attention-based architecture introduced in Attention Is All You Need.', )
 CONFLICT (id) DO  
  label  EXCLUDED.label, category  EXCLUDED.category, tags  EXCLUDED.tags,
  summary  EXCLUDED.summary, community  EXCLUDED.community;



 wiki_edges (source, target, relation, confidence, typed)
 (, , , , )
 CONFLICT (source, target, relation) DO   confidence  EXCLUDED.confidence, typed  EXCLUDED.typed;


 wiki_edges (source, target, relation, confidence, typed)
 (, , , , )
 CONFLICT (source, target, relation) DO   confidence  EXCLUDED.confidence, typed  EXCLUDED.typed;

Write one INSERT ... ON CONFLICT (id) DO UPDATE statement per page (values escaped: single quotes doubled, tags serialized as a JSON array literal cast to jsonb), then one INSERT ... ON CONFLICT (source, target, relation) DO UPDATE per edge. relation stays lowercase here (unlike the uppercased Cypher relationship label) since it's a plain column value, not a schema identifier — this keeps it directly filterable with WHERE relation = 'contradicts' or joinable without case-folding. typed is true only for edges promoted by a relationships: frontmatter entry; plain wikilinks stay false.

Skip pages whose id collides only after the ON CONFLICT clause fires from a prior run — do not attempt to deduplicate synthetic multi-edges (e.g. same source/target with both a wikilink and a typed relation) since the composite primary key (source, target, relation) already keeps them as distinct rows, matching the "typed version wins" merge behavior of graph.json/graph.graphml at the query layer (SELECT * FROM wiki_edges WHERE source=$1 AND target=$2 ORDER BY typed DESC LIMIT 1).

---

3e. graph.html

A self-contained interactive visualization using the vis.js CDN (no local dependencies). The user opens this file in any browser — no server needed.

Build the HTML file by:

1. Generating a JSON array of node objects for vis.js:

{id: "concepts/transformers", label: "Transformer Architecture", color: {background: "#4E79A7"}, size: <degree * 3 + 8>, title: "concepts | #ml #architecture", community: 0}

• Color by community (cycle through: #4E79A7, #F28E2B, #E15759, #76B7B2, #59A14F, #EDC948, #B07AA1, #FF9DA7, #9C755F, #BAB0AC)
• Size by degree (incoming + outgoing link count): size = degree * 3 + 8, capped at 60
• title = tooltip text shown on hover: category, tags, summary (if available)

2. Generating a JSON array of edge objects for vis.js:

// Untyped wikilink
{from: "concepts/transformers", to: "entities/vaswani", dashes: false, width: 1, color: {color: "#666", opacity: 0.6}, title: "wikilink"}
// Typed edge
{from: "concepts/transformers", to: "concepts/lstm", dashes: false, width: 2, color: {color: "#E15759", opacity: 0.8}, label: "contradicts", font: {size: 9, color: "#ccc"}, title: "contradicts"}

• dashes: true for INFERRED edges
• dashes: [4,8] for AMBIGUOUS edges
• Typed edges (typed: true): set width: 2, add a label field showing the type, and apply a type-specific color:

Type	Edge color
extends	#59A14F (green)
implements	#4E79A7 (blue)
contradicts	#E15759 (red)
derived_from	#F28E2B (orange)
uses	#76B7B2 (teal)
replaces	#B07AA1 (purple)
related_to	#BAB0AC (grey — same as untyped)

Untyped wikilink edges keep the existing #666 grey color and no label.

3. Writing the full HTML file:

<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8">
<title>Wiki Knowledge Graph</title>
<script src="https://unpkg.com/vis-network/standalone/umd/vis-network.min.js"></script>
<style>
  * { box-sizing: border-box; margin: 0; padding: 0; }
  body { background: #0f0f1a; color: #e0e0e0; font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif; display: flex; height: 100vh; }
  #graph { flex: 1; }
  #sidebar { width: 260px; background: #1a1a2e; border-left: 1px solid #2a2a4e; padding: 14px; overflow-y: auto; font-size: 13px; }
  #sidebar h3 { : ; : ; : uppercase; : ; :   ; }
   { : ; : ; : ; }
   { : flex; : center; : ; :  ; : ; }
   { : ; : ; : ; : ; }
   { : ; : ; : ; }





  Wiki Knowledge Graph
  Click a node to see details.
  Communities
  
  

Replace /* NODES_JSON */ and /* EDGES_JSON */ with the actual JSON arrays you generated in step 1.

---

Step 3.5: OKF Bundle Export (optional)

Run this step only when the user asks for OKF / a markdown bundle (phrases like "export to OKF", "OKF bundle", "open knowledge format", "export as markdown bundle"). It is additive — the five graph files above are always produced; this writes an extra full-fidelity markdown bundle.

The five graph files are a lossy projection (graph skeleton only). An OKF bundle is the actual page bodies, so an export→wiki-import round-trip through OKF preserves full content, and the bundle drops straight into MkDocs, Notion, Hugo, GitHub's renderer, or any Open Knowledge Format consumer.

Canonical frontmatter mapping (obsidian-wiki ⇄ OKF)

This table is the single source of truth for the mapping; wiki-import references it for the reverse direction.

OKF key	← export from	→ import to	Notes
type (required)	category, title-cased	category (lower-cased)	concepts→Concept, entities→Entity, skills→Skill, references→Reference, synthesis→Synthesis, projects→Project, journal→Journal. OKF requires type; consumers tolerate any string.
title	title	title	Verbatim.
description	summary	summary	Our one-line summary: is exactly OKF's description (used in index.md entries).
tags	tags	tags	Verbatim list. visibility/* system tags pass through unchanged.
timestamp	updated	updated	ISO 8601 both sides.
resource	first sources: entry iff it is an http(s):// URL	—	Optional; omit when no source URL. Most pages describe abstract knowledge and have none.
(extensions)	category, sources, created, relationships, lifecycle, tier, base_confidence, …	preserved verbatim	OKF §4.1 permits arbitrary keys and requires consumers to preserve them. Writing our native keys as OKF extension frontmatter is what makes the round-trip lossless — on import, preserved // are preferred over re-deriving from .

Steps

Reuse the node list from Step 1 (with any active project/visibility filters already applied). Write a directory tree under wiki-export/okf/:

1. One file per in-scope page. For each page, parse its frontmatter, apply the mapping table above to build the OKF frontmatter (required type first, then title, description, tags, timestamp, optional resource, then the preserved extension keys), transform the body links (below), and write to wiki-export/okf/<category>/<slug>.md — same relative path the page has in the vault.

2. Body link transform ([[wikilinks]] → standard markdown links):

   • [[concepts/transformers]] → [<target title>](<file-relative path>.md), e.g. from entities/foo.md a link to concepts/transformers becomes [Transformer Architecture](../concepts/transformers.md). Link text = the target page's title (fall back to the target id if unknown).
   • [[target|display]] → [display](<rel path>.md).
   • Use file-relative paths (../concepts/x.md), never /-absolute — /-rooted links break GitHub rendering. (This matches knowledge-catalog's own production agent.)
   • Compute that relative path from the target file path, not the bare page id: normalize the wikilink target to its page id, append .md, then compute relpath(<target-file>, <source-file-dir>). Never call relpath() on the id before adding .md.
   • This is required for the common "folder note" layout where a page id exists both as a file and as a directory prefix, e.g. projects/social-twitter.md plus projects/social-twitter/.... From projects/social-twitter/concepts/mem0-memory-analysis.md, a link to [[projects/social-twitter]] must export to ../../social-twitter.md, not ...md.
   • Resolve link targets with the same normalization used in Step 1 (lowercase, spaces→hyphens, strip .md). Handle unresolved targets by form, so forward-references survive the round-trip:
     • Resolves to an in-scope page → relative markdown link to it.
     • (contains a , e.g. ) with , and not excluded by an active filter → still emit the relative markdown link. OKF §5.3 treats a missing target as not-yet-written knowledge, and keeping the link makes the user's forward-references lossless on re-import. (Verified on st3ve: dropping these silently deleted real .)

Excluded from the bundle (same as the graph export): _archives/, _raw/, _readouts/, .obsidian/, _insights.md, _meta/*.base, and the vault root index.md (regenerated above).

---

Step 4: Print Summary

Wiki export complete → wiki-export/
  graph.json    — N nodes, M edges (NetworkX node_link format)
  graph.graphml — N nodes, M edges (Gephi / yEd / Cytoscape)
  cypher.txt    — N MERGE nodes + M MERGE relationships (Neo4j)
  postgres.sql  — N upsert rows (wiki_pages) + M upsert rows (wiki_edges) (any Postgres)
  graph.html    — interactive browser visualization (open in any browser)

Append this line only when the OKF bundle was produced (Step 3.5):

  okf/          — OKF v0.1 markdown bundle (N pages, lossless; import via wiki-import)

Append filter notes when active:

  (filtered: project:prismor — 19 of 67 pages)
  (filtered: X of Y pages excluded — visibility/internal, visibility/pii)

Only include lines for filters that were actually applied.

Notes

• Re-running is safe — all output files (and the okf/ bundle) are overwritten on each run
• Broken wikilinks are skipped — only edges to pages that exist in the vault are exported; in the OKF bundle, a wikilink to a missing/filtered page degrades to plain text
• OKF is the lossless format — graph.json reconstructs only stubs on import, while the okf/ bundle preserves full page bodies. Use OKF for vault-to-vault transfer and external markdown tools (MkDocs, Notion, GitHub); use the graph files for analysis tools (Gephi, Neo4j)
• The wiki-export/ directory should be gitignored if the vault is version-controlled — these are derived artifacts
• graph.json is the primary format — the others are derived from it. If a future tool supports graph queries natively, point it at graph.json