Guides

Owned models

Owned models are clients of Inferix. You train and serve them; Inferix observes, routes, watches drift, and promotes revisions. Three families: general-llm, slm-support (cheap CX), slm-apiheal (cheap classify).

Scenario

You have checkpoints on disk and local serve endpoints. You want RouteIQ to send easy CX to slm-support, API classify to slm-apiheal, default reasoning to general-llm, and escalate to a provider when capability or risk says so — with FineForge promoting *-vN after eval gates.

Era path (operate)

Agent task
  → RouteIQ (cheap SLM / general-llm / provider)
  → Inference (owned endpoint or provider)
  → TraceForge + LensAI (model_id labeled)
  → DriftWatch (vs teacher / golden)
  → FineForge (retrain / rollback)
  → RouteIQ rules updated

Register models

models:
  - name: owned/general-llm
    provider: owned
    endpoint: http://general-llm:8080/v1
    revision: general-llm-v3
    capabilities: [chat, tools]
    timeout_ms: 60000

  - name: owned/slm-support
    provider: owned
    endpoint: http://slm-support:8080/v1
    revision: slm-support-v2
    capabilities: [chat]
    labels: { tier: cheap, domain: cx }

  - name: owned/slm-apiheal
    provider: owned
    endpoint: http://slm-apiheal:8080/v1
    revision: slm-apiheal-v2
    capabilities: [chat]          # no heal tools
    labels: { tier: cheap, domain: apiheal }

  - name: provider/strong
    provider: external
    # credentials via env
    capabilities: [chat, tools]

RouteIQ matrix

PathModelUse
Cheap CXowned/slm-supportFAQ, triage, short turns
Cheap classifyowned/slm-apihealSeverity / diff narrate only
General ownedowned/general-llmDefault agent reasoning + tools
Hard pathprovider/strongAmbiguity, risk, outage, heal propose

Details: Routing policies, RouteIQ.

FineForge revisions

  • Version as general-llm-vN, slm-support-vN, slm-apiheal-vN
  • Promote only after eval gates + registry pin
  • Canary → full; rollback restores previous revision
  • Teacher model critiques failures before LoRA/distill

FineForge · Drift & retrain

Ownership

Model cards, eval gates, and model_id live with the model owners. RouteIQ owns cheap/general/hard rules. Agent teams own tools and chaos. Inferix improves how you operate models — it is not required to create them.

Repos

Back to the guides index.

All guides →