# Local workers and models

This overlay targets the separately configured `jobjakt-dev` stack in the primary checkout. It does not replace that checkout's standalone compose file. Production services, databases and model caches are excluded.

The local runtime uses Ollama 0.40.11 and `qwen2.5:7b` (Q4_K_M, Apache-2.0), matching the model name in the repository's production configuration. Exact deployed digest parity has not been established. The frozen official registry manifest is in `model-manifests/qwen2.5-7b.json`; its SHA256 is `845dbda0ea48ed749caafd9e6037047aa19acfcfd82e704d7ca97d631a0b697e`. The image digest and resource limits are pinned in `ollama.compose.yml`. CPU inference is limited to four CPUs and 8GB; cloud processing is disabled. Model files use a dedicated development volume; the API binds only `127.0.0.1:11434`.

Initial local setup, explicitly authorised by the owner on 27 October 2026:

```powershell
rtk proxy docker compose -f dev/ollama.compose.yml up -d --wait
rtk proxy docker compose -f dev/ollama.compose.yml exec ollama ollama pull qwen2.5:7b
rtk proxy python dev/verify_ollama.py --manifest dev/model-manifests/qwen2.5-7b.json --output .artifacts/pre-phase10-ux/local-ai/verification.json
```

Verification checks the actual image, dedicated cache and every model/config/template/license blob against the frozen manifest before inference. A future changed upstream tag must fail verification; review the artifact before updating the pin. The download network remains available for explicit model maintenance; no cloud provider is enabled.

The existing approved DistilBART CPU checkpoint is reused through a read-only mount after all files are checked against `docs/redesign/phase-9/p9-sec-02/full-model/ARTIFACTS.json`. Transformers is offline, so this operation cannot fetch replacement weights. The parser-capable sidecar is also checked against its previously qualified local image ID.

Preflight and apply (run from this worktree; replace paths only with the verified development checkout and approved artifact):

```powershell
rtk proxy python dev/enable_local_workers.py --checkout "E:/Storage/Projects/Active/Job tracker" --bart-artifact "E:/Storage/Projects/Active/Job tracker/.artifacts/p9-sec-02/full-model/artifact" --output .artifacts/pre-phase10-ux/local-ai/worker-preflight.json
rtk proxy python dev/enable_local_workers.py --checkout "E:/Storage/Projects/Active/Job tracker" --bart-artifact "E:/Storage/Projects/Active/Job tracker/.artifacts/p9-sec-02/full-model/artifact" --output .artifacts/pre-phase10-ux/local-ai/workers-applied.json --apply
rtk proxy python dev/check_local_ai.py --output .artifacts/pre-phase10-ux/local-ai/sidecar-smoke.json
```

The apply helper validates project, Development environment, loopback backend ports, development database, current backend image, offline model mount and Mailpit settings. It recreates only the development backend/sidecar without a build or pull, preserving the database and volumes. It enables rules, reminders, saved-search alerts, daily exports, job enrichment and queued AI processing. These features still respect account entitlements, AI consent, configured schedules and available input. Email goes only to Mailpit at `http://localhost:8026`; it is not delivered externally.

This configuration operation does not publish source changes. In particular, the running older backend image cannot use the new `Ai:RequestTimeoutSeconds=240` setting until built from the validated feature source. Its old 30-second limit can interrupt CPU generation. Source builds and application checks have separate evidence.

OAuth/mail synchronisation, billing and external data providers need their own approved credentials and configuration. Enabling workers cannot provide them. CV document import uses deterministic Docling; generative CV-import assistance stays disabled pending its separate source-grounding/accuracy qualification.

## Evidence, 27 October 2026

The local setup and worker apply passed. Independent model-blob verification passed. A real constrained JSON CPU response passed in 20.546 seconds including model loading. Authenticated sidecar checks passed: trained local job summarisation in 2.518 seconds and an Ollama-backed CV rewrite in 40.997 seconds. These checks assert actual nonempty content, not semantic correctness. No production digest parity or CV extraction accuracy is claimed.

Official sources: [Ollama model](https://ollama.com/library/qwen2.5:7b), [frozen registry source](https://registry.ollama.ai/v2/library/qwen2.5/manifests/7b), [Ollama Docker guidance](https://docs.ollama.com/docker), [local/cloud configuration](https://docs.ollama.com/faq).