Get started

Core concepts

Shared vocabulary for operators running Inferix. Use these terms in policies, dashboards, and runbooks.

Glossary

Term	Meaning
Control plane	Inferix process in front of owned models and providers. Observes, routes, watches drift, and kicks retrain — it does not replace your gateway or APM.
Call	One inference request entering /v1/chat/completions (or equivalent). Attributed to a model, agent, tenant, and route decision.
Span / trace	TraceForge unit of work. A trace is a tree of spans: route decision, tool calls, model invoke, downstream HTTP.
Policy	RouteIQ rule set: which model handles a call given intent, agent, tenant, or attributes.
Owned model	Model you host and register in inferix.yaml (e.g. owned/general-llm, owned/slm-support). Inferix routes to its endpoint.
Provider	External inference API registered alongside owned models. Used as primary or fallback per policy.
Route	The chosen model (and optional fallback chain) for a single call after RouteIQ evaluates rules.
Drift	Quality shift vs a baseline or golden set — DriftWatch signals this before user complaints spike.
Golden set	Labeled prompts + expected outputs used by DriftWatch to score quality over a sliding window.
Teacher	Stronger model used by FineForge to generate training targets for a student SLM.
Promote	FineForge action: ship a trained candidate into the live route for a model name.
Rollback	FineForge action: restore the previous promoted revision after a bad ship.

Product roles in one line

LensAI — latency, cost, errors, volume
TraceForge — follow one request across tools and models
RouteIQ — policy routing to owned models or providers
DriftWatch — quality drift alerts
FineForge — retrain, promote, rollback

The operator loop

Observe with LensAI and TraceForge → route with RouteIQ → detect drift with DriftWatch → retrain and promote with FineForge. Every guide in these docs assumes that loop.

Request path, product roles, and data stores.

Architecture →