Mirrored verbatim from utopia-models/knowledge wi/components/ollama.md by the allowlisted publish pipeline. Edit it there, not here.

component: ollama
serves: all
runs-on: novaserver, a system unit running as the `ollama` user
adrs: [6723, 2165]
owns:
  - ollama.service
  - "127.0.0.1:11434"

ollama — local models on novaserver, and nothing that calls them

🔴 RETIRED 2026-10-30 (knowledge#10314, Tyler-approved disk cleanup). ollama.service is stopped and disabled and /usr/share/ollama/.ollama/models (8.2 GB) is deleted; /usr/local/bin/ollama and the unit file remain. curl 127.0.0.1:11434 now refuses. deep-research.sh (nova) fails loud with exit 2. Restoring it means systemctl enable --now ollama plus re-pulling models. The rest of this page describes the pre-retirement state.

WHAT IT IS

ollama.service, bound to 127.0.0.1:11434: small local models for cheap, offline-capable work — classification, embeddings, quick transforms — on a box with no GPU. If it stops, only one thing notices: deep-research.sh in nova's brain (~/.claude/scripts/, outside this repo), which defaults to BACKEND=ollama. No script, workflow, hook or unit in this repo calls it.

🔴 It is not a dispatch backend and cannot become one by configuration. ADR-6723 makes the fleet 100% Claude Code on Claude subscriptions; dispatch.sh passes the literal TASK_BACKEND=claude and there is no per-master backend field for this to appear in. ADR-2165 already disqualified it on measurement, not preference — "no GPU, 7 tok/s, 5-22 min prefill per call, impossible at 350/day".

HOW IT WORKS TODAY

flowchart LR
    S["ollama.service (system,<br/>user `ollama`)"] --> D["/usr/local/bin/ollama serve"]
    D --> P["127.0.0.1:11434<br/>LOOPBACK ONLY"]
    P -.->|"connection refused —<br/>correct, not an outage"| T["anything on the tailnet"]
    P --> M["models pulled into<br/>~ollama/.ollama/models"]
    X["the dispatch chain"] -.->|"TASK_BACKEND=claude<br/>(ADR-6723)"| Y["never reaches here"]

One unit, one process, no repo artifact and no drop-in. Everything about it is read live — the commands are in DETERMINE CURRENT STATE.

The model list is off-git and this doc deliberately does not state it — a pull or a prune changes it with no commit anywhere, and the previous doc's list outlived two of its own facts. What is worth knowing structurally is the size class: these are 1.5B-7B instruct and embedding models, and that is a property of the box (no GPU), not of the list.

Loopback is a design choice, not an oversight. ss -tlnp | grep 11434 shows LISTEN 127.0.0.1:11434, so a caller on the tailnet gets a connection refusal. That is correct behaviour. Any fleet use would be from a process on novaserver itself, or through an explicit forward — never by rebinding the port.

Where its cost lands. It shares RAM with the SigNoz containers and four master harnesses on one 12-core box, and a loaded 7B model is the largest single thing it can add. Measure before blaming or excusing it:

systemctl show ollama -p MemoryCurrent --value    # bytes; 0 or [not set] means nothing is loaded

BOUND VALUES

• 🔴 ollama — no units: binding, so nothing proves its ExecStart target still exists (#4742)

No file, port list or model count is bindable here — the unit is a distro install and the models are on-disk state. The registry binding above is the only in-repo claim this component makes, and it is the one that would catch the unit being removed.

DETERMINE CURRENT STATE

systemctl is-active ollama
curl -s -m5 http://127.0.0.1:11434/api/tags | python3 -c \
  "import json,sys;[print(m['name'],m['details'].get('parameter_size'),m['size']) for m in json.load(sys.stdin)['models']]"
systemctl show ollama -p MemoryCurrent --value    # bytes; 0 or [not set] means nothing is loaded

The model list is off-git and changes with a pull or a prune, committing nothing — so it is read, never recalled.

KNOWN BROKEN

• 🔴 No caller in this repo. grep -rln "11434\|ollama" --include='*.sh' --include='*.py' --include='*.mjs' --include='*.yml' _infra/ .github/ returns only _infra/ansible/roles/system_units/defaults/main.yml, which manages the unit itself (2026-10-31). Under ADR-6723 that is the expected state; recorded here so the next reader does not spend a run looking for the consumer. 🔴 Per the plan's own rule, low traffic is a deletion candidate, never a verdict — and its one caller is nova's deep-research.sh, outside the repo's infra tree.
• ⚠ A 1.5B-7B model used for reasoning-shaped work fails silently. The output is confident, fluent and wrong, and it arrives with a 200. That is the failure mode to design against if anything ever does call this.
• ⚠ The old reference doc pins four model names and a date ("confirmed 2026-08-01"). Model names are exactly the class of off-git fact that rots without announcing itself; this doc replaces the list with GET /api/tags.