Deterministic showcase
Multimodal Product Listing Generator
A catalogue row and a product image go in; a validated listing, a deterministic quality report and a set of claim candidates come out. What is worth looking at here is not the copy — it is everything the pipeline does around the model so a response is never treated as a trusted database record.
Try it
Choose an image and fill in the catalogue fields, then generate. The image is validated and previewed in your browser and never uploaded; only the catalogue fields reach the route, which returns a deterministic fixture put through the same validation, quality and claim layers as a live response.
JPG, PNG or WEBP, up to 5 MB. Previewed locally only.
Deterministic showcase mode. No API key is used, no live model call is made, and the uploaded image is not sent anywhere.
What runs in this demo
Validation boundary
The catalogue row is validated before anything else runs. A blank field or a non-positive price is refused at the boundary, not discovered downstream.
Strict schema
Output is checked against a strict schema with unknown fields rejected, so a malformed or hallucinated response shape can never be silently accepted.
Deterministic quality
Title length, description word count, feature count and keyword count are computed in code. The model is never asked to grade its own output.
The canonical implementation is a Python package. This page is a TypeScript presentation adapter that ports the deterministic layers faithfully; the multimodal prompt construction, injected transport with selective retries, defensive parsing of fenced JSON, safe logging and the review workflow all live in the repository.
Evidence, and its boundary
514
Tests passing
98.26%
Core-package coverage
90%
CI coverage floor
3 / 3
Deterministic evaluation checks
What these numbers do not say
A 3/3 deterministic pass rate is a property of the fixture, not a model result. It shows the validation layer accepts conforming output and the pipeline runs end to end offline. It says nothing about how a live vision model performs.
- Live model quality: not measured
- Human review: not measured
- Hallucination rate: not measured
- User or business impact: not measured
Reproduce it
python -m product_generator --mock # deterministic run, no key, no network
python -m product_generator.evaluation # deterministic evaluation, 3/3 checks
pytest --cov # 514 tests, 98.26% coverage