# AI Live Pod CAG/DoD Architecture

📘 Actiq CAG + DoD Architecture — Public Documentation Draft

Actiq combines on-device AI, distillation-on-demand (DoD), and a decentralized knowledge layer (CAG) to create a fast, private, and evolving intelligence system that learns from user interaction and scales through distributed memory.

This document outlines the architecture, purpose, and benefits of Cache-Augmented Generation (CAG) and on-demand distillation within the Actiq ecosystem.

CAG replaces traditional retrieval-augmented generation (RAG) with a pre-computed, immutable cache of distilled knowledge blocks. These blocks are embedded directly into prompts or fine-tuning datasets for tiny LLMs, enabling:

• Faster response times (no retrieval delay)
• Higher consistency and relevance
• Seamless integration with local and decentralized storage

Unlike RAG, which relies on a live search through a vector database at inference time, CAG uses previously curated and verified QA pairs (or micro-knowledge units) stored in a globally accessible layer. These are:

• Distilled from trusted LLMs (e.g., GPT-4, Claude)
• Filtered for relevance, safety, and clarity
• Stored on decentralized infrastructure like Arweave
• Signed and tagged with metadata for future reuse

CAG functions as a shared knowledge memory that many devices and agents can draw from without needing repeated cloud queries or retrieval cycles. This dramatically reduces latency, cost, and reliance on central infrastructure.

           ┌─────────────────────┐
           │   Cloud LLM (GPT)   │
           └────────┬────────────┘
                    │
              [DoD Query Trigger]
                    ↓
           ┌────────┴────────────┐
           │   Distilled Answer  │
           │  + Metadata + Tags  │
           └────────┬────────────┘
                    ↓
          ┌─────────▼────────────┐
          │  Arweave CAG Layer   │
          │ (Immutable QA Blocks)│
          └───────┬──────────────┘
                  ↓
    ┌─────────────▼───────────────┐
    │ Actiq Edge Device (Tiny AI) │
    │  ┌───────────────┐          │
    │  │ Local Model   │◄─────────┐
    │  └───────────────┘          │
    │  ┌───────────────┐          │
    │  │ Local RAG DB  │          │
    │  └───────────────┘          │
    └──────────┬──────────────────┘
               ↓
    [User Prompt / Real-Time Feedback]

DoD is the process of selectively querying large cloud models (e.g., GPT-4, Claude, Gemini) only when the on-device model lacks confidence. The returned response is then:

1. Returned to the user in real time
2. Saved to a permanent knowledge cache (CAG)
3. Used to improve the local model via distillation or tuning

This feedback loop lets Actiq devices improve continuously without over-relying on cloud inference.

Actiq stores verified QA blocks on Arweave — a permanent, decentralized data layer — making these distilled insights:

• Publicly auditable and accessible
• Cost-efficient (one-time cost ~$0.000006 per QA block)
• Syncable across devices without duplication

The cache layer acts as a collective intelligence memory, with contributions from many users and devices. Each entry is:

• Hashed and signed by its origin device
• Labeled with source LLM, tags, and optional confidence rating
• Optionally rated by the community or private to device

Access can be:

• Open/public (global common knowledge)
• Permissioned (team, project, product)
• Private (device-only cache)

[User prompt] → [Tiny model answers]
     ↓
[If confidence low]
     ↓
→ [Query Cloud LLM (DoD)]
     ↓
→ [Answer returned + saved to Arweave (CAG)]
     ↓
→ [Local distillation / LoRA update]
     ↓
→ [Future prompts use CAG context directly]

• Private by design: Personal context stays local; shared knowledge is anonymized and distilled
• Efficient: Reduces cloud cost and latency by over 80%
• Decentralized: Each user contributes to and benefits from a shared intelligence layer
• Scalable: CAG can support millions of edge nodes without bloated cloud infra

{
  "prompt": "How does CAG work?",
  "source": "GPT-4",
  "confidence": 0.92,
  "response": "CAG uses cached QA blocks to generate answers without retrieval.",
  "tags": ["CAG", "distillation"],
  "hash": "0xA9321F...",
  "public": true
}

• Personalized sport coaching — real-time body-aware AI trainers
• Cognitive health & neuroplasticity — mental + physical coordination
• Embodied AI assistants — human-like home or retail guides
• Agent networks (DePIN) — collective reasoning across devices

• Fine-grained access permissions / token gates
• Community-weighted voting for CAG relevance
• LoRA adapters market for niche distillation modules
• Real-time syncing with Arweave Lite nodes

This architecture represents a shift in AI deployment: from centralized, reactive tools to proactive, adaptive, and distributed intelligence — embodied in every Actiq device.

For questions or integrations, contact: [email protected]

➡️ Visit: docs.actiq.xyz