Examples
Real calls and their results. Run any in the playground; for whole tasks, see the cookbook.
Vision input
Read a photographed receipt
A crumpled receipt photo in; typed totals, items and a computed discount out.
Read a CV into typed records
A CV image into a typed candidate, one record per job, and every skill.
Find a dog head in a photo
A photo of a dog in; a bounding box around its head.
Find every dog head in a photo
A photo of four dogs in; a box around each head.
Image question answering with visual reference
Has any bear caught a fish? The answer with its confidence, and a box around the bear and the fish that show it.
Complex workflows in one API call
Sort an email, then ask what that kind needs
An invoice gets its amount, a meeting request its date, a job application its role.
Classify a ticket, refund only for billing, and reply
Category and urgency side by side, the refund only for billing, then a reply from both.
Discover a new category
Sort an expense into your categories, or name a new one when none fits.
Hierarchical classification
Pick a department, then a category inside it, in one call.
List who was there and what they will do
An array returns as many names as the meeting notes hold.
Audit an expense claim
Read every line of a claim, then check it, list the rules broken and decide, in one call.
Look up an order and ask an agent mid-call
Stop for your order database, and for a lost order an agent's refund decision, then reply.
Thinking when a question is difficult
Better probabilities and confidence
Route a support ticket
Route a ticket with a probability for every queue, so unsure ones go to a person.
Refund only when sure
Ask for the refund amount only when the ticket is billing and the model is sure.
Roll a fair die
Averaging over option orders brings a fair die back to one in six for every face.
Screen two identical résumés
Two identical résumés: the fair answer is 50/50, whichever is listed first.