Knowledge Distiller v3.1

Extract Knowledge, Not Text.

An AI agent skill that distills any document into structured knowledge graphs with temporal awareness. Atomic concepts, explicit relationships, time context, zero redundancy.

Dual Output: .knowledge.md (human-readable, Obsidian-compatible) + .knowledge.json (machine-readable, JSON-LD graph)

What It Does

What it does NOT do	What it DOES
1:1 transcript of the document	Identify and describe concepts atomically
Copy-paste with Markdown formatting	Make relationships between concepts explicit
Rebuild the table of contents	Eliminate redundancy (same idea = one block)
	Build temporal knowledge structure

Key Features

• Knowledge Graph — Rich nodes with definitions, statements, relevance. Weighted, typed edges. JSON-LD with schema.org context.
• Temporal Dimension (v3.1) — Three temporal layers: source period, validity range, distillation date. Every node, fact, and chunk carries time context.
• Hierarchical Clusters — Concepts grouped into thematic clusters with a concept map overview.
• Embedding-Ready Chunks — Clean-text chunks without Markdown syntax. Each starts with a natural-language time reference.
• Per-Block Confidence — high / medium / low based on evidence in the source document.
• Obsidian Wikilinks — [[concept]] syntax for cross-references. Drop into your vault.
• Any File Format — PDF, DOCX, PPTX, XLSX, CSV, JSON, HTML, TXT, images, URLs.

Example Output

The examples/ directory contains a full distillation of Anthropic's Claude Skills Guide:

• claude-skills-guide.knowledge.md — 13 knowledge blocks, 7 clusters, 13 facts with temporal context
• claude-skills-guide.knowledge.json — JSON-LD graph with 13 nodes, 15 edges, 13 embedding-ready chunks

Temporal Dimension (v3.1)

When you distill documents across years, knowledge without a time axis creates contradictions. "Revenue: 28B" and "Revenue: 34B" aren't conflicting facts — they're a time series.

Layer	Fields	Purpose
Source	source_date, source_period	When was the knowledge published?
Validity	valid_from, valid_until	When does this information apply?
Distillation	distilled	When was the extraction performed?

Time formats: ISO 8601 with flexible granularity — 2024-12-31, 2024-Q4, FY2024, 2020/2024

Temporal confidence: explicit (date stated in document) / inferred (derived from context) / unknown

Graph Schema

Eight relationship types:

Type	Meaning	Symbol
uses	Dependency	→
enables	Causality	→
based-on	Foundation	→
part-of	Composition	→
tension	Trade-off / Contradiction	↔
replaces	Supersession	→
extends	Extension	→
example-of	Instantiation	→

Repository Structure

knowledge-distiller/
├── SKILL.md                 # The skill definition (v3.1)
├── README.md
├── LICENSE
├── index.html               # Landing page (GitHub Pages)
├── style.css
├── base.css
├── docs/                    # Landing page (copy)
│   ├── index.html
│   ├── style.css
│   └── base.css
└── examples/
    ├── claude-skills-guide.knowledge.md
    └── claude-skills-guide.knowledge.json

Works With

Tool	How
Obsidian	Drop .knowledge.md files into your vault. Wikilinks create a navigable graph.
Neo4j / Gephi	Import .knowledge.json nodes and edges directly.
OpenAI / Cohere Embeddings	Use the chunks[].embedding_text field — clean text, no syntax artifacts.
RAG Pipelines	Each chunk is self-contained with time context. Ready for LangChain / LlamaIndex.

Version History

Version	Date	Changes
v3.1	16 March 2026	Temporal dimension: three time layers, per-node temporal objects, time-stamped facts and chunks
v3.0	16 March 2026	Dual output (.md + .json), rich graph nodes, embedding-ready chunks, JSON-LD
v2.0	15 March 2026	Hierarchical clusters, machine-readable graph, per-block confidence
v1.0	15 March 2026	Initial release: 6-phase workflow, atomic blocks, Obsidian wikilinks

License

MIT

---

Created with Perplexity Computer