Turning Liquid Thoughts into Solid Knowledge.
An AI-powered interactive workspace where concepts become draggable, fracturable, mergeable nodes. Built with DeepSeek V4 and vanilla JavaScript — no frameworks, no GPU, no Wasm required.
Every concept is decomposed through a strict ontological hierarchy:
| Level | Name | Example |
|---|---|---|
| L1 | Category/Field | Anatomy, Transportation |
| L2 | System/Entity | Human Skeleton, Car |
| L3 | Macro-Parts/Sub-systems | Axial Skeleton, Engine |
| L4 | Micro-Components | Skull, Piston |
| L5 | Materials/Tissues | Bone Tissue, Steel |
| L6 | Chemical/Atomic | Calcium, Carbon |
The AI assesses the input's LOD and decomposes exactly one level deeper — never skipping from L2 (Skeleton) to L6 (Calcium).
No forced component counts. If a system naturally has 2 parts, you get 2. If it has many, you get the most critical 3-8. Quality and structural accuracy over arbitrary quotas.
- 0ms Knowledge Cache: Every unique keyword is cached server-side. Repeated queries return instantly without API calls.
- Model Escalation: If
deepseek-v4-flash(LITE) returns empty, the system auto-escalates todeepseek-chat(STANDARD) with JSON mode enabled. No manual intervention needed. - Three-Tier Routing: LITE (flash) → STANDARD (chat) → ULTRA (pro) — complexity-based auto-selection.
Every AI response undergoes real-time Quality Control. A separate critique agent scores the output on accuracy, creativity, depth, and relevance. If the score is below 7, the response is regenerated with improvement guidance.
Ultra-lite DOM architecture with translucent panels, backdrop blur, and subtle glow effects. All nodes are draggable <div> elements — no canvas rendering overhead.
Three AI personas (Scientist, Philosopher, Artist) independently decompose a concept, then a Synthesizer combines their perspectives into a final result. Alt+double-click to trigger.
- Semantic Memory: Persistent knowledge graph tracking all relationships
- Analogical Reasoning:
A:B :: C:?— AI completes analogies from your workspace - Counterfactual Reasoning: "What if X didn't exist?" / "What is the extreme of X?"
- Proactive Curiosity: AI suggests interesting merges when you're idle
- Tension Detection: Identifies dialectical oppositions with wavy SVG lines
- Deep Fracture: Recursive multi-level decomposition (Shift+double-click)
- Workspace Persistence: Auto-save to localStorage, export/import as JSON
- Real-Time Collaboration: WebSocket relay for multi-user workspaces
- Embedding Similarity: Vector-based semantic clustering (press 'C')
- Ontology Hierarchy: IS-A taxonomy chains with LCA merge context (press 'H')
- Tool-Using Agent: Wikipedia, dictionary, calculator tools via ReAct loop
- Hover Descriptions: AI-generated definitions with cached async loading
| Layer | Technology |
|---|---|
| AI Backend | DeepSeek V4 (Flash / Chat / Pro) |
| Server | Node.js HTTP + WebSocket relay |
| Frontend | Vanilla JS (ES Modules), Preact+htm HUD |
| Styling | CSS Glassmorphism, SVG overlays |
| Storage | localStorage + server-side memory cache |
| Validation | Custom pre-commit hook (scripts/validate.js) |
- Node.js v18+
- DeepSeek API Key (free tier available)
git clone https://github.com/akaradje/Liquid-State-Engine.git
cd Liquid-State-Engine
npm installCopy the environment template and add your API key:
cp .env.example .env
# Edit .env: DEEPSEEK_API_KEY=sk-your-key-herenpm run serve
# Open http://localhost:8080npm run collab
# WebSocket relay on ws://localhost:8081| Gesture | Action |
|---|---|
| Double-click workspace | Create a new node |
| Drag node | Move freely |
| Double-click node | AI fracture (decompose) |
| Shift + Double-click | Deep recursive fracture |
| Alt + Double-click | Multi-agent debate |
| Drag A onto B | AI merge (synthesize) |
| Right-click / long-press | Context menu (fracture, invert, amplify, delete) |
| Hover (500ms) | AI-generated description + metadata tooltip |
| Press 'S' | Save workspace |
| Press 'L' | Load workspace |
| Press 'C' | Auto-cluster by similarity |
| Press 'H' | Hierarchy view |
| Press 'R' | Toggle relationship lines |
| Press 'M' | Toggle audio |
| Press 'F' | Toggle reflection |
├── scripts/
│ ├── serve.js # HTTP server + 15 AI endpoints
│ ├── collab-server.js # WebSocket relay
│ ├── validate.js # Pre-commit validator
│ └── test-ai.js # 19-test AI suite
├── web/
│ ├── index.html # Entry point
│ ├── main.js # Frontend orchestrator
│ ├── style.css # Glassmorphism + animations
│ ├── ai-enrich.js # AI enrichment client
│ ├── ai-memory.js # Semantic knowledge graph
│ ├── ai-suggest.js # Auto-suggestion client
│ ├── audio.js # Procedural sound effects
│ ├── collab.js # WebSocket collaboration
│ ├── curiosity.js # Proactive AI suggestions
│ ├── meta-learner.js # User preference learning
│ ├── ontology.js # IS-A hierarchy tree
│ ├── particles.js # Fracture/merge effects
│ ├── persistence.js # Save/load workspace
│ ├── relations-viz.js # SVG relationship lines
│ ├── semantic-space.js # Embedding similarity
│ ├── tension-detector.js # Dialectical opposition
│ ├── trails.js # Energy beam canvas
│ └── hud/
│ ├── app.js # Preact HUD
│ └── style.css # HUD panel styles
├── .env.example # API key template
├── CLAUDE.md # AI coding rules
└── package.json
MIT
Built for the era of Liquid Information — where concepts flow, merge, and crystallize into knowledge.