Aeion AI Architecture
How Aeion's AI layer fits together — a single routing engine in front of 57 provider adapters, with routing-strategy selection, automatic same-modality provider fallback, and cost recording on every request; AeionClaw skill calls add per-tenant budget enforcement on top.
Engine Capabilities
Routing engine
One entry point sits in front of every provider — resolving routing strategy and same-modality provider fallback before any call goes out. AeionClaw skill calls add a per-tenant budget check on top of this same engine.
Provider registry
57 provider adapters register into a single registry; the router picks the right one per model name automatically.
Cost engine
Built-in pricing tables compute cost per call, per modality — tokens, images, audio seconds, video seconds.
Usage tracking
Every call is logged with tokens, cost, latency, and status for per-tenant reporting.
Data model
Three collections back the module — provider configs, usage logs, and dashboard configs — all queryable through the standard API.
Model registry
Auto-refreshes daily so new provider models show up without a deploy.
Vector store
Built-in retrieval-augmented generation (RAG) support.
Rate limiting
Redis-backed, per-tenant and per-provider.
Response caching
Exact-match caching today; semantic caching on the roadmap.
Tool use
A unified tool-use / function-calling abstraction across providers.
Structured output
Schema-validated LLM responses — pass a schema, get typed JSON back.
Conversation state
Multi-turn context management built in.
Prompt management
Templating and versioning for production prompts.
Neural Bus integration
The platform's intelligent rule engine runs on top of this layer.
System agent
A meta-agent that uses the engine to power platform-wide automation.
Adapters, Registry, Engine — the Three-Layer Pattern
Same three-layer pattern used throughout Aeion OS: adapters, a registry, and one engine in front of both.
Routing Strategies
Model selection is driven by a routingStrategy preference, per-call selectable.
Fallback Chain — Resilience Built-In
Real-world LLM operations: providers go down. Anthropic rate-limits at 4PM. OpenAI 503s. Google throttles. The engine falls through to another active provider for the same modality without app-layer retry logic.
End-to-End Request Flow
What happens between your call and the response coming back.
The Provider Contract
Every one of the 57 adapters honors the same contract — that's what makes adding a new provider mechanical rather than a rewrite: