Visual Product Search
A customer sends a photo; the workflow reads any printed code first and only calls a vision model when it has to.
Problem
Customers photograph a product tag and ask 'do you have this' instead of typing a code or name. Matching that by hand meant a staff member squinting at a low-res photo.
Before
Photos landed in the same inbox as text messages and waited for someone to visually match them against the catalogue.
After
Product codes are read and matched automatically. Photos without a readable code are identified by colour, pattern and type using a vision model — only when needed.
Interactive workflow
Click through the actual node graph
The cost-routing decision
This is the part worth explaining rather than hiding: most product photos in this business already contain a printed code. OCR is cheap and fast, so it runs first. Only when OCR comes back empty — no readable code — does the workflow escalate to a vision model call, which is slower and costs more per request. Routing the cheap path first is what keeps a busy page's AI bill proportional to actual ambiguous cases, not to total photo volume.
Flow
An exact code match skips the AI step entirely — it's a direct catalogue lookup. Only the no-code path aggregates the catalogue and asks the vision model to identify the item by what it looks like.
Impact
Photo enquiries get an answer as fast as text ones, without a proportional rise in the vision-API bill.