🧠 Semantic Brain

Semantic Brain is an autonomous knowledge graph generator and explorer. It uses LLMs (via node-llama-cpp) to dynamically expand conceptual landscapes, evaluating relationships across multiple semantic dimensions and visualizing them in a real-time 3D environment.

Demo

Here's a simple demo with few pre-generated graphs:

https://karolrybak.github.io/semantic_brain/?demo=true

🛠 Setup & Requirements

1. Model Configuration

Configure your model paths in config.json in the project root. You must use models compatible with GBNF JSON grammar (e.g., Qwen-2.5-Instruct or Llama-3-Instruct).

{
  "modelPaths": {
    "small": "path/to/small-model.gguf",
    "medium": "path/to/medium-model.gguf",
    "large": "path/to/large-model.gguf"
  },
  "selectedSize": "medium",
  "loadOnStartup": true
}

2. Installation

bun install

3. Running the Application

You need to run the frontend dev server and the WebSocket logic server in separate terminals:

Terminal A (Frontend):

bun run dev

Terminal B (AI Server):

bun run server

🖥 User Interface Guide

The 3D Workspace

• Navigation: Click and drag to rotate, scroll to zoom, right-click to pan.
• Nodes: Represent concepts. Sphere size indicates general relevance; color represents health/status.
• Links: Visual connections between concepts. AI-generated links (Blue), User-created (White), and Bridge links (Emerald).

Graph Explorer (Left Panel)

Appears when a node is selected:

• Description: AI-generated summary of the concept.
• Aspect Relevance: Real-time breakdown of how the concept scores against your defined semantic dimensions.
• Discovery Actions:
  • Discover New: AI brainstorms new related concepts based on your allowed relations.
  • Link Existing: AI looks for logical connections between the selected node and other concepts already in your graph.
• Refresh: Use the icon in the header to re-generate descriptions or re-evaluate aspects.

Control Center (Top Right)

• 🔙 Back to List: Exit the current graph and return to the workspace selector.
• 🔄 Re-layout: Re-heat the physics simulation to organize the nodes.
• 🧠 Brain Panel: Monitor model status, check VRAM usage, switch model sizes, or manually Load/Unload the engine.
• 🔗 Relations Management: Toggle which ontological types (Causal, Structural, Logical, etc.) the AI is allowed to use for discovery.
• ⚙️ Config: Adjust AI creativity, visual link opacity, label size, and toggle emoji visibility.

Semantic Bar (Top Center)

• Manage the "Aspects" (dimensions) of your analysis. Toggle specific aspects to see the graph physically rearrange itself—nodes will gravitate towards the center based on their relevance to the active aspects.

⚠️ Troubleshooting

JSON Grammar Issues

If you see errors regarding createGrammarForJsonSchema, your version of node-llama-cpp might need a manual recompile to support the latest features.

1. Follow the official build instructions.
2. Or use the built-in shortcut:

bun run recompile

This will force a clean rebuild of the native binaries and close the process. Afterward, start the server normally.

---

Note: This project is optimized for Bun. Compatibility with standard Node.js is currently untested.