Skip to content

EXECUTION_MODES.md

Latest commit

 

History

History
55 lines (34 loc) · 2.79 KB

EXECUTION_MODES.md

File metadata and controls

55 lines (34 loc) · 2.79 KB
Raw

AuraRouter: How Your Tasks are Solved

AuraRouter is not just a "pass-through" for your AI prompts. It uses an intelligent loop to figure out the best way to answer your question. This approach, called Federated Mixture-of-Experts (FMoE), ensures you get high-quality results without the high cost of big cloud models.

1. The Standard Path: Intent -> Plan -> Execute (IPE)

Most AI tasks follow a simple "Input -> Output" path. AuraRouter adds two smart steps in the middle to save you time and money.

Step 1: The Classifier (Triage)

Instead of sending your prompt directly to a large model, a tiny, ultra-fast model first "triages" your request. It determines:

  • Intent: Is this a simple question, a coding task, or a complex analysis?
  • Complexity: On a scale of 1-10, how hard is this?

Step 2: The Planner

If your task is complex (e.g., "Build a full website with a database"), AuraRouter doesn't just start typing. A reasoning model creates a Plan—a list of logical steps needed to finish the job.

Step 3: The Worker

Finally, the task is sent to a Specialist. If it's a coding task, it goes to a model trained specifically for code. If it's a creative task, it goes to a writing specialist.

Why do this? Traditional methods send every "Hello" to a $20/month cloud giant. AuraRouter sends the "Hello" to a free local model and saves the cloud giant for the tasks that actually need it.

2. Advanced Reasoning: AuraMonologue

For the hardest tasks (Complexity 8-10), AuraRouter uses AuraMonologue. This is like having a "committee of experts" in your computer.

  • The Generator: Creates the initial answer.
  • The Critic: Reviews the answer for errors, missing details, or bad logic.
  • The Refiner: Takes the Critic's feedback and fixes the answer.

The loop continues until the Critic is satisfied. This ensures that even local, smaller models can produce "genius-level" results by checking their own work.

3. Speed vs. Quality: Speculative Decoding

If you want the speed of a small model but the quality of a large one, AuraRouter uses Speculative Decoding.

  1. A small "Drafter" model quickly writes a response.
  2. A large "Verifier" model checks the draft in the background.
  3. If the draft is correct, it's shown to you instantly. If there's a mistake, the verifier fixes it before you even notice.

4. Local First, Cloud Last

AuraRouter is designed for Sovereignty.

  • Privacy: It checks your prompts for sensitive data. If it finds any, it forces the task to stay on your local hardware.
  • Cost: It tracks every cent you save. By using local models for 80% of your tasks, you drastically reduce your monthly AI bill.
  • Control: You are the "Architect." You decide which models are your favorites and which roles they should play.