context

name

context

description

Ingest and structure knowledge from any source: PDFs, web pages, YouTube videos (with transcription), links, and notes. Everything is stored in OKF (Open Knowledge Format) with a mind graph of concepts and relations. TRIGGER when: user says 'guarda esto', 'remember this', 'ingest', 'extract concepts', 'build knowledge graph', 'add to my context', 'remember this PDF/link/video/note', or asks about knowledge management, OKF format, concept graphs, or mind maps. SKIP: user wants to search existing knowledge (use vector search), needs a database query, or is asking about RAG implementation details.

Context — Knowledge Ingestion Skill

Ingest knowledge from any source into OKF (Open Knowledge Format), build a mind graph of concepts, and store it automatically.

How It Works

1. Ingest: Fetch/extrac content from any source
2. Analyze: Extract concepts and relations (using this Claude session)
3. Store: Save to mind graph (automatically, no asking)
4. Report: Show what was added

OKF Schema

interface OKFDocument {
  id: string;                    // okf_<timestamp>_<random>
  type: 'web' | 'pdf' | 'video' | 'note' | 'link';
  title: string;
  content: string;
  url?: string;
  metadata: {
    created: string;
    source: 'web' | 'pdf' | 'yt' | 'manual';
    language?: string;
    duration?: number;
    authors?: string[];
    tags: string[];
  };
  concepts: Array<{ id: string, label: string, importance: number, mentions: number }>;
  relations: Array<{ from: string, to: string, type: string, weight: number }>;
}

Mind Graph

Stored at ~/.divagator/graph.json. Auto-persisted on every ingest.

mkdir -p ~/.divagator
touch ~/.divagator/graph.json

Format:

{
  "nodes": [
    { "id": "concept_1", "label": "CachedNetworkImage", "importance": 0.9, "mentions": 8, "sources": ["okf_123_abc"] }
  ],
  "edges": [
    { "from": "concept_1", "to": "concept_2", "type": "extends", "weight": 0.8, "sources": ["okf_123_abc"] }
  ]
}

How to Execute (Automatic Flow)

Step 1: Identify Source Type

Web URL: https://... PDF file: /path/to/file.pdf YouTube: https://youtube.com/... or yt:... Note: Any raw text (no URL prefix)

Step 2: Ingest Content

Web Page:

curl -s -L -A "Mozilla/5.0" "<URL>" | \
  sed 's/<script[^>]*>.*<\/script>//g' | \
  sed 's/<style[^>]*>.*<\/style>//g' | \
  sed 's/<[^>]*>/ /g' | \
  tr -s ' \n' ' ' | \
  cut -c1-15000

PDF:

pdftotext /path/to/file.pdf - 2>/dev/null | cut -c1-15000

YouTube Video:

mkdir -p /tmp/divagator-yt

# Download audio
yt-dlp -x --audio-format mp3 \
  -o "/tmp/divagator-yt/%(id)s.%(ext)s" \
  "<YOUTUBE_URL>"

# Transcribe with lightning-whisper-mlx
VIDEO_ID=$(yt-dlp --get-id "<YOUTUBE_URL>")
AUDIO_FILE="/tmp/divagator-yt/${VIDEO_ID}.mp3"

python3 -c "
from lightning_whisper_mlx import LightningWhisperMLX
model = LightningWhisperMLX(model='small', batch_size=16)
result = model.transcribe('$AUDIO_FILE')
print(result['text'])
"

Note: User provided text directly.

Step 3: Extract Title

• Web/PDF: Extract from content or filename
• YouTube: Use video title from yt-dlp
• Note: User provided or generate from first line

Step 4: Extract Concepts (Claude Analysis)

Analyze the content within this conversation to extract:

• Concepts: Named entities, key terms, important phrases
• Importance: 0.0-1.0 based on mention frequency and centrality
• Relations: Connect concepts, assign type (relates, extends, contradicts, unknown)
• Weight: 0.0-1.0 based on connection strength

Step 5: Build OKFDocument

{
  "id": "okf_<timestamp>_<random>",
  "type": "video|web|pdf|note|link",
  "title": "<title>",
  "content": "<content>",
  "url": "<url if applicable>",
  "metadata": {
    "created": "<ISO timestamp>",
    "source": "yt|web|pdf|manual",
    "tags": []
  },
  "concepts": [...],
  "relations": [...]
}

Step 6: Update Mind Graph (Automatic)

1. Load existing graph from ~/.divagator/graph.json
2. Add new concepts (merge if concept label already exists, update importance/mentions)
3. Add new relations (merge if relation already exists, update weight)
4. Save back to ~/.divagator/graph.json

Step 7: Report

✓ Added to context:

**Source**: [title](url)
**Type**: web/pdf/video/note
**Concepts**: 9 extracted
**Relations**: 4 new connections

**Mind Graph**: 24 concepts, 31 relations now in ~/.divagator/graph.json

Invocation Examples

/context ingest https://example.com/article
/context add this PDF: /path/to/paper.pdf
/context remember this YouTube: https://youtube.com/watch?v=...
/context add a note: "La memoria se consolida durante el sueño REM"

Query Commands

/context graph — show full mind graph summary
/context concepts about <topic> — find concepts related to topic
/context stats — show graph statistics

Environment

Variable	Default	Description
YT_OUTPUT	/tmp/divagator-yt	YouTube audio temp

Notes

• No API key needed: Concept extraction uses this Claude conversation
• Auto-persist: Graph saved to ~/.divagator/graph.json after every ingest
• Merge behavior: Existing concepts are merged by label (highest importance wins)
• YouTube: Requires yt-dlp and lightning-whisper-mlx (pip install lightning-whisper-mlx)