PRD: Personal Knowledge Graph with AI-Assisted Curation

A guide for setting up a private semantic knowledge base with automated maintenance

Overview

This document describes how to build a personal knowledge graph using Wikibase, integrated with OpenClaw for AI-assisted curation. The system enables "write-ahead linking" where you casually reference topics in your notes/tasks, and automation handles entity creation, research, and quality assurance.

Target Audience: Another OpenClaw instance or technical user wanting to replicate this setup.

Architecture

Components

Component	Purpose	Self-Hosted?
Wikibase	Structured knowledge storage with SPARQL queries	Yes (Docker/K8s)
Fizzy	Task management with cards and comments	Yes (Rails app)
OpenClaw	AI assistant for automation and curation	Yes
Obsidian	Long-form notes and documentation	Local app

Data Flow

User writes [[wb:Topic]] in Fizzy card
         ↓
Cron job scans for placeholders (every 6h)
         ↓
Search Wikibase for existing entity
         ↓
If not found: Create stub + research card
         ↓
Enrichment job adds properties (every 8h)
         ↓
Review job checks quality (daily)
         ↓
Rich, queryable knowledge graph

Setup Instructions

1. Deploy Wikibase

Use the official Wikibase Docker Suite or deploy to Kubernetes:

# Example K8s deployment (simplified)
apiVersion: apps/v1
kind: Deployment
metadata:
  name: wikibase
spec:
  containers:
  - name: wikibase
    image: wikibase/wikibase:latest
    env:
    - name: MW_ADMIN_PASS
      valueFrom:
        secretKeyRef:
          name: wikibase-secrets
          key: admin-password

Key Configuration:

Enable the Wikibase API
Set up admin credentials (store in 1Password or similar)
Configure SPARQL endpoint for queries

2. Create Base Schema

Define foundational types and properties:

Type Items (Classes):

QID	Label	Description
Q1	Person	Human being
Q2	Company	Business organization
Q3	Project	Work project or initiative
Q4	Place	Geographic location
Q5	Event	Occurrence in time
Q6	Concept	Abstract idea or category

Core Properties:

PID	Label	Type	Description
P1	instance of	Item	Type/class of entity
P2	date of birth	Time	Birth date
P6	start date	Time	When something began
P20	website	URL	Web presence
P24	creator	Item	Who made this
P25	uses technology	Item	Tech stack

3. Set Up Link Reconciliation

Script: wikibase-link-reconcile.sh

This script scans your task system for [[wb:Entity Name]] patterns and creates Wikibase entities.

#!/usr/bin/env bash
# Core logic (simplified)

# 1. Authenticate with Wikibase
wikibase_login

# 2. Get all cards/notes with [[wb:...]] patterns
patterns=$(grep -oP '\[\[wb:[^\]]+\]\]' "$content")

# 3. For each pattern:
for pattern in $patterns; do
    entity_name="${pattern//\[\[wb:/}"
    entity_name="${entity_name//\]\]/}"
    
    # Search Wikibase
    qid=$(search_wikibase "$entity_name")
    
    if [[ -z "$qid" ]]; then
        # Create stub
        qid=$(create_stub_entity "$entity_name")
        
        # Create research card for enrichment
        create_research_card "$entity_name" "$qid"
    fi
    
    # Replace placeholder with real URL
    replace_placeholder "$pattern" "$WIKIBASE_URL/wiki/Item:$qid"
done

Cron Schedule: Every 6 hours

OpenClaw Cron Job:

{
  "name": "Wikibase Link Reconciliation",
  "schedule": {"kind": "every", "everyMs": 21600000},
  "payload": {
    "kind": "systemEvent",
    "text": "Run wikibase link reconciliation: /path/to/script.sh"
  },
  "sessionTarget": "main"
}

4. Set Up Entity Enrichment

Script: wikibase-enrich-entities.sh

Finds stub entities and triggers AI research to enrich them.

# Core logic (simplified)

# 1. Query recently created/modified entities
entities=$(get_recent_entities)

# 2. Check if each is a stub
for qid in $entities; do
    if is_stub "$qid"; then
        echo "$qid|$label" >> /tmp/stubs_to_enrich.txt
    fi
done

# 3. Output list for AI session to process

OpenClaw Cron Job (Isolated Session):

{
  "name": "Wikibase Entity Enrichment",
  "schedule": {"kind": "every", "everyMs": 28800000},
  "payload": {
    "kind": "agentTurn",
    "message": "Execute enrichment script, then for each stub in /tmp/stubs_to_enrich.txt: research the topic via web search and add Wikibase properties."
  },
  "sessionTarget": "isolated"
}

5. Set Up Weekly Review

Script: wikibase-weekly-review.sh

Daily QA pass on entities from the last 7 days.

Quality Checks:

Has label and description
Description > 20 characters
Has instance_of (P1) claim
Has at least 2 claims total

Output: Markdown report in Obsidian

Cron Schedule: Daily at 2am local time

Usage Patterns

Write-Ahead Linking

When writing in Fizzy (or any integrated system), use the placeholder syntax:

Working on integrating [[wb:Home Assistant]] with [[wb:OpenClaw]].
The setup runs on [[wb:Proxmox VE]] using [[wb:Kubernetes]].

Within 6 hours, each becomes a real link. If the entity doesn't exist, a stub is created and a research card queued.

Filtering Examples

Skip patterns that look like documentation examples:

# In extract_wb_links function
clean_text=$(echo "$text" | sed 's/`[^`]*`//g')  # Remove code blocks
patterns=$(echo "$clean_text" | grep -v -E '\[\[wb:(Entity Name|\.\.\.)\]\]')

Stub Detection Heuristics

An entity is considered a stub if:

Description < 50 characters, OR
Fewer than 2 claims, OR
Missing instance_of (P1)

Prompts for AI Sessions

Enrichment Session Prompt

Execute /path/to/wikibase-enrich-entities.sh to scan for stubs.

For each stub in /tmp/wikibase_pending_enrichment.txt:
1. Read the current label and description
2. Search the web for authoritative information
3. Add relevant Wikibase properties:
   - P1 (instance of) - classify the entity type
   - Relevant relationships based on entity type
4. Write a comprehensive description (2-3 sentences)
5. Update the entity via Wikibase API

Review Session Prompt

Run /path/to/wikibase-weekly-review.sh

Review the generated report at /path/to/reports/wikibase-review/YYYY-MM-DD.md

For entities with critical issues:
- Create Fizzy cards for items needing human judgment
- Auto-fix issues where confident (e.g., add missing P1 for obvious types)

Integration Points

Obsidian Sync (Future)

Sync Wikibase claims to Obsidian frontmatter:

---
wikibase: Q35
instance_of: knowledge graph
creator: Nathan Broadbent
start_date: 2026-01
---

Natural Language Queries (Future)

Expose SPARQL via natural language:

"Who do I know in Auckland?" → Query persons with residence = Auckland
"What projects use Kubernetes?" → Query items with uses_technology = Kubernetes

File Locations

File	Purpose
`/scripts/wikibase-link-reconcile.sh`	Link reconciliation script
`/scripts/wikibase-enrich-entities.sh`	Enrichment scanner
`/scripts/wikibase-weekly-review.sh`	QA review script
`/obsidian/Knowledge Base/Schema.md`	Property and type definitions
`/obsidian/Knowledge Base/Current State.md`	Status overview
`/obsidian/reports/wikibase-review/`	Daily QA reports

Troubleshooting

Authentication Issues

Wikibase uses MediaWiki login API
Store credentials in 1Password: op://vault/wikibase/admin_password
Cookie file: /tmp/wb_cookies.txt

Entity Creation Fails

Check CSRF token is valid (refresh after login)
Ensure JSON is properly escaped
Verify API permissions for WikibaseAdmin user

Stubs Not Being Detected

Check lookback window (default 24h for enrichment)
Verify recentchanges API is returning data
Namespace 120 = Item namespace in Wikibase

Success Metrics

Entity count grows steadily over time
Stub ratio decreases (more entities fully enriched)
Weekly review pass rate improves
Link placeholders get resolved within 6 hours

This PRD enables any OpenClaw instance to replicate the Broadbent Knowledge Graph setup. Adapt paths, credentials, and schedules to your environment.

ndbroadbent/personal-knowledge-graph.md

Select an option

No results found