Local Notes Assistant
Summarise synthetic operations notes through a bounded local model call, with validated JSON and a model-free fixture path.
What the demonstration does
Reads a bounded notes file, asks the local model for a summary and actions, checks the response schema, and optionally writes a local JSON result. It does not send messages or update a task system. The included notes are synthetic.
Run the fixture path
Download and extract the archive, then run from its root:
python3 -m labs.local_notes.main --fixture
python3 -m unittest discover -s labs/tests -vFixture mode uses a canned response and labels it explicitly. It exercises application handling without making an inference-performance claim.
Connect Ollama
Install and start Ollama following the official platform guide, then:
ollama pull qwen3:4b
python3 -m labs.local_notes.main --output runs/notes-result.jsonThe helper calls the loopback API with a timeout, bounded context/output, and JSON mode. It rejects unexpected fields and invalid types. A schema-valid response still needs factual review before it becomes an action.
Verification and boundaries
Model-free tests cover input limits, malformed responses, JSON parsing, and schema validation. Live Ollama inference was not tested in the initial release. The separate Delivery-Note Model Lab uses MLX; its timings do not establish this model’s performance.
The source archive includes the program, synthetic input, shared helper, tests, and a README. Python’s standard library is sufficient for the fixture path.
Extend carefully
Before using real notes, agree on the data boundary and operating setup. For a batch process, add a checksum-based input manifest, bounded retries, a rejection queue, and reviewed output. Keep task creation or distribution as a separately approved step.