Problem
How does a camera observation become a structured list of ingredient names, then a retrieved document, then a recommendation — without pretending a local library is GPT-4o?
What it is / is not
It is a local-first reading of Pic2Plate: name schema, coverage score, retrieve-evaluate-refine loop, and transcribed paper figures. It is not the official implementation, does not ship model weights, and does not generate recipes from a hidden API key.
Architecture
image
↓
vision → structured names (and stop)
↓
retrieve → score → rewrite query → retrieve
↓
recommendation document
Evidence
- Paper: doi:10.3390/s25020449
- Official notebooks: leonardcl/pic2plate
- This repo cites the paper and keeps interfaces local.