wiki: update AI-Stack
chaib
parent
b486036a62
commit
e462fac844
1 changed files with 0 additions and 104 deletions
104
AI-Stack.md
104
AI-Stack.md
|
|
@ -1,104 +0,0 @@
|
|||
# AI Stack
|
||||
|
||||
The AI layer is the reasoning core of Atlas OS: a durable decision spine, an autonomous-but-gated execution Brain, and a cost-tiered model router that always reaches for the cheapest capable model first. This page covers how those pieces fit together and where the rough edges still are.
|
||||
|
||||
For where these services surface in the desktop, see [[Atlas-OS]]. For hosting and isolation, see [[Infrastructure]] and [[Security]].
|
||||
|
||||
---
|
||||
|
||||
## At a glance
|
||||
|
||||
| Component | What it is | Where it lives |
|
||||
|---|---|---|
|
||||
| Kernel spine | `kernel.sqlite` — the durable record of decisions and events | Local SQLite spine |
|
||||
| Atlas Brain | Reviewed-execution autonomous core | Reasoning layer over the spine |
|
||||
| Policy gate | Human-approval checkpoint for sensitive actions | In front of every Brain action |
|
||||
| LLM router | Cost-tiered model selector | `localhost:8888` |
|
||||
| Headroom proxy | Context compression layer | In front of model calls |
|
||||
|
||||
Surfaced in Atlas OS under the **AI** category: Atlas Chat (Open WebUI), AI Gateway, Mindmap, and Leeragent.
|
||||
|
||||
---
|
||||
|
||||
## The spine — `kernel.sqlite`
|
||||
|
||||
Everything the Brain does is anchored to a single durable record. `kernel.sqlite` holds roughly **458 decisions** and **3,415 events**, giving the system an auditable memory of what was decided, when, and what happened around it.
|
||||
|
||||
The spine is the source of truth: decisions are recorded as first-class objects, and events accumulate as an append-style log around them. This is what lets autonomous execution stay reviewable — there is always a record to inspect rather than an opaque chain of model calls.
|
||||
|
||||
---
|
||||
|
||||
## The Brain — reviewed execution
|
||||
|
||||
**Atlas Brain** is the autonomous core, built on a *reviewed-execution* model. It can reason, plan, and act against the spine, but it is not unconditionally free to do so.
|
||||
|
||||
A **policy gate** sits in front of every action. It **always requires human approval** for three classes of action:
|
||||
|
||||
- **Money** — anything financial.
|
||||
- **External communications** — anything that leaves the system to a third party.
|
||||
- **Self-modification** — anything that changes Atlas itself.
|
||||
|
||||
Actions outside those classes can proceed under reviewed execution; the three gated classes never auto-execute. This is a hard rule, not a tunable setting — the gate is the safety boundary that makes autonomy acceptable.
|
||||
|
||||
---
|
||||
|
||||
## The router — cost-tiered model selection
|
||||
|
||||
The LLM router at **`localhost:8888`** picks the cheapest model that can handle a task, and only escalates when it must. The ordering is deliberate: local models cost nothing and stay private, free cloud tiers cost nothing but leave the box, and paid Claude is the last resort.
|
||||
|
||||
| Tier | Models | Cost profile |
|
||||
|---|---|---|
|
||||
| **1 — Local** | LLaMA 1B, hermes3, Qwen Coder, Qwen 7B | Zero cost, fully on-box |
|
||||
| **2 — Free cloud** | Groq `llama-3.3-70b`, Cerebras, OpenRouter | Zero cost, off-box |
|
||||
| **3 — Paid** | Claude | Used last, only when needed |
|
||||
|
||||
The principle: never burn paid tokens on work a free or local model can do. Classification, routing, copy, and summaries stay local; harder reasoning escalates only as far as it has to.
|
||||
|
||||
---
|
||||
|
||||
## Headroom — context compression
|
||||
|
||||
The **Headroom proxy** sits in front of model calls and compresses context before it reaches the model. This keeps prompts within working limits and reduces the cost and latency of each call — which in turn makes the cheaper tiers viable for more tasks than they otherwise would be.
|
||||
|
||||
---
|
||||
|
||||
## How a request flows
|
||||
|
||||
```mermaid
|
||||
%%{init:{'theme':'base','themeVariables':{'primaryColor':'#0c1428','primaryTextColor':'#cfe2ff','primaryBorderColor':'#1F6FE0','lineColor':'#4EA0FF'}}}%%
|
||||
flowchart TD
|
||||
REQ[Incoming task] --> BRAIN[Atlas Brain<br/>reviewed execution]
|
||||
BRAIN --> SPINE[(kernel.sqlite<br/>decisions + events)]
|
||||
BRAIN --> GATE{Policy gate}
|
||||
|
||||
GATE -->|money / external-comms / self-modify| HUMAN[Human approval required]
|
||||
GATE -->|everything else| HEADROOM[Headroom proxy<br/>context compression]
|
||||
|
||||
HUMAN -->|approved| HEADROOM
|
||||
|
||||
HEADROOM --> ROUTER{LLM router<br/>localhost:8888}
|
||||
ROUTER -->|cheapest capable| LOCAL[Local models<br/>LLaMA 1B / hermes3 / Qwen]
|
||||
ROUTER -->|escalate| FREE[Free cloud<br/>Groq / Cerebras / OpenRouter]
|
||||
ROUTER -->|last resort| CLAUDE[Claude paid]
|
||||
|
||||
LOCAL --> SPINE
|
||||
FREE --> SPINE
|
||||
CLAUDE --> SPINE
|
||||
```
|
||||
|
||||
The Brain records its reasoning to the spine, passes through the policy gate, and — once cleared — compresses context via Headroom before the router selects a model from the cheapest capable tier. Results flow back into the spine.
|
||||
|
||||
---
|
||||
|
||||
## Known gap — cross-device chat history
|
||||
|
||||
Chat-history consolidation across devices is **not yet solved**. Conversation history is not currently unified into a single cross-device view, and reconciling it is a pending follow-up rather than a shipped feature. Treat per-device history as authoritative on its own device until this is addressed.
|
||||
|
||||
---
|
||||
|
||||
## Related pages
|
||||
|
||||
- [[Atlas-OS]] — the desktop where AI apps surface
|
||||
- [[Infrastructure]] — hosting, Caddy, containers
|
||||
- [[Security]] — EU-sovereign hosting, secrets, MFA
|
||||
- [[Fleet-and-Mesh]] — NATS mesh and node-to-node messaging
|
||||
Loading…
Reference in a new issue