Obsidian AI Knowledge Management Guide

💡TL;DR

Notemd turns LLM-powered reading into persistent knowledge: wiki-links connect concepts, concept notes create a retrievable graph, research brings the web into your vault, translation breaks language barriers, diagrams make structure visible, and workflows chain it all into one click. This guide covers the full pipeline — from raw notes to a connected, visual, multilingual knowledge base.

Why AI Knowledge Management?

Traditional note-taking produces flat files. Even with manual wiki-links, most notes stay disconnected. Notemd uses LLMs to automate the connection layer:

• LLMs read your content and identify what matters — terms, methods, people, theories
• Links are inserted automatically at each concept occurrence, not buried in "see also"
• Concept notes are generated as standalone retrievable files
• Research enriches notes with web-sourced context
• Diagrams make structure visible — mind maps, flowcharts, data charts from same content

The result: a knowledge graph that grows with every note you process, not just when you remember to add links.

The Full Pipeline

Each step is independent. Use one or all. The most impactful sequence: Add Links → Concept Notes → Diagrams.

---

1. Wiki-Links: Making Connections Explicit

Wiki-links are the backbone of a knowledge graph. Notemd uses an LLM to:

1. Read your note content (split into chunks for long documents)
2. Identify core concepts — prioritizing specific, technical terms over generic nouns
3. Insert [[wiki-links]] at each occurrence
4. Suppress synonyms so "ML" and "Machine Learning" don't create separate nodes

When to Use

• Every note >100 words — shorter notes yield few concepts
• Research papers, technical docs, meeting notes — rich in domain-specific terms
• After content is stable — don't repeatedly process drafts

Key Settings

Setting	Recommended	Why
addLinksProvider	DeepSeek or GPT-4o-mini	Good accuracy at low cost
Synonym suppression	On	Prevents duplicate nodes
Context window	Paragraph	Balance of accuracy and cost

→ Wiki-Links deep dive

---

2. Concept Notes: Retrievable Knowledge Nodes

Wiki-links connect ideas inline, but concept notes make each idea independently retrievable. Each concept gets its own .md file:

# Machine Learning

## Linked From
- [[My Research Notes]]
- [[Neural Networks Explained]]

The Extraction Process

The LLM prompt is highly structured:

• Normalize to singular form
• Prefer multi-word concepts over single words ("Dielectric Relaxation" not "Relaxation")
• Skip references/bibliography sections
• Output as CONCEPT: lines for deterministic parsing

Concepts are deduplicated across chunks via Set<string>. LLM errors on individual chunks don't abort the operation.

Backlinks

When enabled, each concept note tracks which source notes mention it. Obsidian's native backlink panel also shows reverse connections.

Deduplication

Notemd's 4-step dedup engine catches:

1. Exact matches — case-insensitive filename comparison
2. Plural forms — "Models.md" vs "Model.md"
3. Symbol normalization — "A-B.md" vs "A B.md"
4. Single-word containment — "ML.md" flagged when "Machine Learning.md" exists

Key Settings

Setting	Recommended	Why
conceptNoteFolder	concepts/ or 🧠 concepts/	Keeps vault organized
extractConceptsAddBacklink	On	Enables reverse lookup
extractConceptsMinimalTemplate	Off	Full template with Linked From
Per-task model	DeepSeek	Concept extraction doesn't need expensive models
Synonym suppression	On	Same setting affects both linking and extraction

→ Concept Notes deep dive

---

3. Research: Bringing the Web In

Notemd integrates web search into your note-taking workflow:

1. Query construction — your note title or selection becomes a search query
2. Web search — Tavily (recommended, API key required) or DuckDuckGo (free, no key)
3. LLM summarization — search results are condensed into a relevant summary
4. Append to note — summary added at cursor position or as a new section

When to Use

• Before processing a new topic — get web context first
• When a concept note needs enrichment — research then add links
• For literature reviews — batch-research a folder of notes

Key Settings

Setting	Recommended	Why
researchProvider	GPT-4o or Claude	Research needs higher quality summarization
Search service	Tavily	Better relevance, configurable depth
maxResearchContentTokens	4000	Balance between depth and cost

→ Research deep dive

---

4. Translation: Breaking Language Barriers

Notemd translates notes using your configured LLM — not a dedicated translation API. This means:

• Context-aware translations — the LLM understands the full document, not sentence-by-sentence
• Technical term handling — "gradient descent" stays as "梯度下降" not "坡度向下"
• Batch support — translate an entire folder of notes in one operation
• Per-task model — use Gemini Flash for translation (fast, cheap, multilingual)

Language Support

Notemd itself supports 21 UI languages. The translation target language is configurable per-task. Common pairs: EN↔ZH, EN↔JA, EN↔KO, EN↔DE, EN↔FR, EN↔ES.

→ Translation deep dive

---

5. Diagrams: Making Structure Visible

Notemd's diagram pipeline is spec-first: the LLM produces a structured DiagramSpec JSON, then adapters translate it into the target format. This produces more reliable output than asking the LLM for raw Mermaid syntax.

Intent Detection

Notemd infers the best diagram type from content:

• Tables with numbers → data chart (Vega-Lite)
• Client/server vocabulary → sequence diagram (Mermaid)
• Entity/primary key → ER diagram (Mermaid)
• Step/process flow → flowchart (Mermaid)
• Concept map keywords → JSON Canvas (Obsidian native)
• Default → mind map (Mermaid)

Rendering Chain

Primary target → fallback → fallback → HTML. If Mermaid syntax fails, it retries once with error context to the LLM, then falls back to a minimal diagram.

Key Settings

Setting	Recommended	Why
enableExperimentalDiagramPipeline	On	Better quality via spec-first
experimentalDiagramCompatibilityMode	best-fit	Native target per intent
summarizeToMermaidProvider	GPT-4o or Claude	Diagram specs need spatial reasoning
autoMermaidFixAfterGenerate	On	Catches LLM syntax errors automatically
Local knowledge augmentation	On for domain-specific	Improves accuracy with vault context

→ Diagrams deep dive

---

6. Workflows: One-Click Automation

Workflows chain multiple tasks into a single sidebar button. The DSL format is:

task1 | task2 | task3

Example: addLinks | extractConcepts | generateDiagram — process a note from raw text to a fully connected, visual knowledge node in one click.

Recommended Workflows

Workflow	Chain	Use Case
Full Process	addLinks | extractConcepts | generateDiagram	New notes
Research First	research | addLinks	Unfamiliar topics
Polyglot	translate | addLinks	Multilingual notes
Diagram Only	generateDiagram	Quick visualization

→ Workflows deep dive

---

7. LLM Providers: 36 Options from Cloud to Local

Notemd supports 36 providers across 4 transport types. Key groups:

• International cloud: OpenAI, Anthropic, Google, Mistral, xAI
• China cloud: DeepSeek, Qwen, Doubao, Moonshot, GLM, Baidu, SiliconFlow
• Gateways: OpenRouter, GitHub Models, Hugging Face, Vercel
• Local: Ollama, LMStudio, OVMS — no API key, no data leaves your machine

Per-Task Model Strategy

The most cost-effective setup uses cheap models for simple tasks and powerful models for complex ones:

extractConcepts  → DeepSeek (fast, cheap, accurate enough)
addLinks          → DeepSeek or GPT-4o-mini
research          → GPT-4o or Claude (needs quality)
generateDiagram   → GPT-4o or Claude (needs spatial reasoning)
translate         → Gemini Flash (fast, multilingual)

→ LLM Providers overview

---

Getting Started Checklist

1. Install Notemd — Community Plugins (recommended) or manual
2. Configure a provider — DeepSeek (easiest), OpenAI, or Ollama (free)
3. Process your first note — right-click → "Process file (add links)"
4. Set concept folder — Settings → Notemd → Output → Concept Folder
5. Extract concepts — run "Extract concepts" on the same note
6. Generate a diagram — run "Generate diagram" to visualize the connections
7. Create a workflow — chain the above into a one-click button

Recommended Configurations

Student (Budget)

Provider: DeepSeek (free tier available)
Concept extraction: DeepSeek
Research: DuckDuckGo (free) + DeepSeek
Diagrams: Off (or legacy Mermaid)
Workflows: addLinks | extractConcepts

Researcher (Quality)

Provider: GPT-4o (primary)
Concept extraction: DeepSeek (cost savings)
Research: GPT-4o + Tavily
Diagrams: best-fit mode, GPT-4o
Workflows: research | addLinks | extractConcepts | generateDiagram

Privacy-First (Local Only)

Provider: Ollama (llama3 or qwen2.5:7b)
All tasks: Ollama
Research: DuckDuckGo (free, no API key)
Diagrams: legacy Mermaid mode

Bilingual (ZH + EN)

Primary: DeepSeek (Chinese queries)
Translation: Google Gemini Flash
Research: Tavily + DeepSeek (Chinese search context)
Language output: per-task (extractConceptsLanguage: zh-CN)

---

Common Patterns

Pattern: Process a Research Paper

1. Import PDF content (or paste)
2. Research — get web context on the topic
3. Add Links — identify and link key concepts
4. Extract Concepts — create standalone notes
5. Generate Diagram — visualize the paper's structure

Pattern: Daily Note Enrichment

1. Write daily note
2. Add Links — connects today's ideas to existing concepts
3. Concept notes auto-update with backlinks

Pattern: Literature Review

1. Create folder with papers/notes
2. Batch Add Links — process entire folder
3. Deduplicate Concepts — clean up near-duplicate notes
4. Generate Diagram — mind map of the entire literature

---

Notemd is open source (MIT) and works with Obsidian 0.15.0+ on all platforms. Install now or view on GitHub.