The classifier backend

Cicero has one model that writes replies. Some decisions are not replies: was that utterance addressed to me, what kind of request is this, should this route somewhere else. Those run per utterance, and running a reply-tier model on every utterance is not affordable — on a single-GPU box it is not even schedulable alongside STT and TTS.

The classifier role is a second, smaller model held apart for exactly that work. It is optional, off by default, and nothing today requires it.

Configuration

Same shape as an llm: section:

classifier:
  backend: llama-cpp          # any backend the llm role supports
  host: 127.0.0.1
  port: 8093                  # not 8090 — that is web voice's default
  model: your-small-model

A ~2B-class instruct model is the intended size. It must be warm — a model that loads on demand defeats the purpose, because the latency lands in the path of every utterance. Cicero sends one throwaway completion to the classifier at startup so the model is resident before the first real one arrives; a health check alone would not do it, since none of the backends load a model on a health probe (ollama's /api/tags lists what is available, it does not load anything). The warmup is best-effort and runs in the background — startup does not wait on it, and a classifier that fails to answer is not fatal because every caller already falls back.

Remote and local-managed both work. host pointing at another machine makes it a remote endpoint Cicero only talks to; a local backend that Cicero can launch is started and stopped with the daemon like any other managed provider.

What absence means

Leaving the section out means the role is unconfigured, and every feature that depends on it stays off and logs the reason.

It does not silently fall back to the reply model. That fallback is the failure this role exists to prevent: it would look like a working feature while quietly costing a full reply per utterance, and on a VRAM-tight box it would contend with the STT/TTS seat. There is no implicit default either — unlike llm, an absent classifier section resolves to nothing at all.

A configured but unsupported backend is an error at startup, not a fallback. The message names the classifier (Unknown classifier backend '<name>') rather than the reply model, so it is clear which of the two sections is wrong, and cicero doctor points at classifier.backend for the same reason.

Failure behavior

The classifier is never a required primary. If it is configured and cannot start, the daemon logs a warning and carries on — a per-utterance helper being down must not stop a daemon that can still hold a conversation. Features that need it decline rather than guess.

cicero doctor reports a configured-but-unreachable classifier as a failing check, exactly as it does for an unreachable reply model — the role was asked for and is not there. An unconfigured one is silent, since absence is a valid choice rather than a problem. cicero status shows a Classifier row only when the role is configured.

VRAM

A resident classifier competes with the STT and TTS seats on a single-GPU box. Measure the resident cost on your hardware before enabling anything that depends on it — this is the reason the role ships optional rather than on.