Deep Dive · Ingesting Knowledge

From File to Knowledge Node

How Context Memory turns 15+ file formats into structured knowledge.

01
15+ Formats — What Goes In

Upload a file and tell Claude to ingest it. The skill automatically picks the best extraction method: PDF→pypdf, Word→pandoc, Excel→pandas, PowerPoint→python-pptx.

📕
PDF
Text per page. OCR fallback for scans.
📝
Word
Headings = section boundaries.
📊
Excel
Columns, types, samples per sheet.
🎬
PowerPoint
Slide text + speaker notes.
🔧
JSON/CSV
Structure analysis, schema, stats.
💻
Code & more
Markdown, HTML, EPUB, YAML, images…
02
Splitting — When to Use What

The question: Store a 50-page PDF as one node or many? Depends on the document.

single → One node per file. For short docs.
sections → Per heading. For specs and manuals.
pages → Per PDF page. When no headings exist.
slides → Per slide. For presentations.
sheets → Per Excel sheet. For thematically separated tabs.
03
Duplicate Detection

If you ingest the same file twice, the skill warns you. It compares title overlap and tag similarity, then offers: save anyway, update existing, or cancel.

Claude: ⚠️ Similar node found:
  → arch-003: Stanford Study (similarity: 0.85)
  Save anyway, update existing, or cancel?
04
Tips for Better Results

Add tags to make knowledge searchable. Add time context to prevent contradictions. Set relevance to help search ranking. And: link immediately when ingesting — saves connecting later.

🏷️
Add Tags
"…with tags: research, llm, stanford"
🕐
Add Time Context
"This is from June 2025"
🔗
Link Immediately
"This confirms the Stanford study"
✅
Quality Check
"How good is my knowledge base?"