feat(processor): make Ollama model + request timeout configurable

[agu2347]

Problem

The Ollama model was hardcoded in processing-service/processor.py:126:

response = _ollama_client.chat(model='llama3.2:1b', messages=[...])

OLLAMA_HOST was already configurable via env var, but the model and any client options were not - changing the model required editing source.

Fix

• Added OLLAMA_MODEL (default llama3.2:1b, unchanged) and OLLAMA_REQUEST_TIMEOUT (default 30 seconds) env vars, following the exact os.getenv() pattern already used for OLLAMA_HOST.
• _ollama_client.chat(...) now uses model=OLLAMA_MODEL instead of the literal.
• OLLAMA_REQUEST_TIMEOUT is passed to ollama.Client(host=OLLAMA_HOST, timeout=OLLAMA_REQUEST_TIMEOUT). The ollama client (0.4.7, as pinned in requirements.txt) forwards **kwargs straight to its underlying httpx.Client, which does accept timeout - verified directly against the installed package's constructor signature, not assumed.
• Documented both vars in .env.example, docker-compose.yml, and k8s/06-processing-service.yml, next to the existing OLLAMA_HOST entries.

Tests

Added two tests to processing-service/test_processor.py:

• test_run_ai_analysis_uses_configured_model - asserts the chat() call receives model=OLLAMA_MODEL, not a literal.
• test_ollama_model_env_var_overrides_default - asserts the default is unchanged (llama3.2:1b), then reconfigures the model and asserts run_ai_analysis picks up the new value with no code change.

Note on the second test: I initially wrote it with importlib.reload(processor) to simulate a fresh process picking up a changed OLLAMA_MODEL env var, but that reruns every module-level statement, including the prometheus_client.Counter/Histogram registrations - which raises ValueError: Duplicated timeseries in CollectorRegistry on the second import. Switched to monkeypatch.setattr(processor, "OLLAMA_MODEL", ...) on the already-imported module instead, which exercises the same code path in run_ai_analysis (a plain module-global lookup) without the reload side effect, and auto-reverts after the test.

Verification

• pytest processing-service/test_processor.py -v - 25 passed (23 pre-existing + 2 new), run against the pinned dependency versions from requirements.txt (kafka-python 2.0.6, elasticsearch 8.17.0, drain3 0.9.11, scikit-learn 1.6.1, prometheus-client 0.21.1, ollama 0.4.7, joblib 1.4.2, regex 2024.11.6).
• ruff check against the project's actual pyproject.toml config (select = ["E","W","F","I","N"], line-length 100) - clean on both changed files.
• docker-compose.yml and k8s/06-processing-service.yml parse as valid YAML (yaml.safe_load_all), and I checked tests/integration/test_k8s_manifests.py - its structural checks (apiVersion/kind/metadata/resource requests) are unaffected by adding two more env: entries.

Acceptance criteria (from the issue)

• Setting OLLAMA_MODEL changes the model used without code changes
• Default unchanged (llama3.2:1b)
• .env.example / k8s / docker-compose updated

Closes #31

[github-actions]

Hey @agu2347! 🌟

Welcome to Sentinel — and thank you so much for jumping in with feat(processor): make Ollama model + request timeout configurable! We genuinely appreciate you taking the time to contribute.

Here's what happens next:

• Someone on the team will review your changes in the next day or two
• If anything needs tweaking, we'll let you know — no pressure, we're happy to help
• Once it looks good, we'll get it merged in

In the meantime, if you have questions about your PR or the project, just reply here. We're around!

Thanks again for being part of this. 🙌

[Uday9909]

Yo this is a clean one. I verified the ollama.Client timeout kwarg actually exists in the pinned ollama 0.4.7, the env vars are documented in all three places (.env.example, docker-compose, k8s manifest), and CI is green. Fully closes #31.

One thing. test_ollama_model_env_var_overrides_default in test_processor.py does not actually test the env var. It monkeypatches processor.OLLAMA_MODEL directly, so the os.getenv line in processor.py is never exercised. If someone replaced os.getenv with the hardcoded literal, that test would still pass. Either rename it to something like test_run_ai_analysis_uses_module_ollama_model, or actually set the env var and reload the module. You already found reload blows up on the duplicate prometheus registrations, so renaming is the honest move unless you split the os.getenv read into a tiny helper worth testing.

Small nit: OLLAMA_REQUEST_TIMEOUT does float(os.getenv(...)) at import, so a non-numeric value in the env crashes the whole processor before it starts. Not blocking, but a try/except with a log would make that failure obvious instead of a cryptic traceback.

[Uday9909]

@agu2347 another thing before opening a pr on an issue , please first get it assigned to yourself!