Docs

Use cases and domains

Where Fern models are already in use, and more domains where a sentence-to-action model fits.

A Fern model fits wherever people repeat the same kinds of request against an API: find these records, change those, start this action. This page lists what has been built and suggests other domains to try.

Three kinds of action

Most requests fall into one of three kinds. The intent bar handles each one differently.

KindWhat it doesExample
FilterFinds records, without changing anything"orders from Germany over €200 this month"
Bulk editChanges a set of records in one step"close every duplicate ticket created today"
ActionDoes one thing"refund order 1042 but not the shipping"

What has been built

WooCommerce stores. Hoobert is a WordPress plugin that adds a command bar to the WooCommerce admin. A shop manager types "refund order 1042 but not the shipping" and Hoobert runs the right REST API call, as that manager. Studio's WooCommerce template starts a project with the same 28 store tools and over 5,500 training pairs.

Calendars and CRMs. The playground has models trained on a calendar tool set ("book a call next Monday at 10:30 with Raj") and a CRM tool set ("move deal 401 to proposal"). They are the quickest way to see a trained model answer.

The homepage shows more sentences and the calls they become, for issue trackers, store admin, support desks, billing, CRMs, calendars and inventory.

Domains to try

These are illustrations, not templates. Each shows one sentence, the call a model trained on that domain would return, and the tools a project for it might start with.

HR and people operations

"approve Sam's leave for next Friday"

{ "name": "approve_leave", "arguments": { "employee": "Sam", "date": "next Friday" } }

An action. A starting tool set: list_leave_requests, approve_leave, reject_leave, list_employees, update_employee. Filters such as "everyone on leave next week" make up most of the traffic in tools like these.

Logistics and fleet

"reroute shipment 4471 to the Leeds depot"

{ "name": "reroute_shipment", "arguments": { "shipment_id": "4471", "destination": "Leeds depot" } }

An action. A starting tool set: list_shipments (with status, origin, destination and date filters), reroute_shipment, assign_driver, mark_delayed. Give location arguments a fixed list of depots or warehouses where you can, so the model chooses a valid one.

Property management

"list vacant units under $1,800 in Brooklyn"

{ "name": "list_units", "arguments": { "status": "vacant", "max_rent": 1800, "area": "Brooklyn" } }

A filter. A starting tool set: list_units, list_tenants, create_maintenance_request, update_rent. Ranges such as "under $1,800" work best as separate minimum and maximum arguments.

Cloud and DevOps consoles

"scale the api service to 4 replicas in staging"

{ "name": "scale_service", "arguments": { "service": "api", "replicas": 4, "environment": "staging" } }

An action. A starting tool set: list_services, scale_service, restart_service, list_deployments, rollback_deployment. For changes like these, show the call to the user before running it, as the intent bar does, and keep production changes behind a confirmation in your own code.

Learning platforms

"give section B two more days on assignment 3"

{ "name": "extend_deadline", "arguments": { "assignment": 3, "section": "B", "days": 2 } }

A bulk edit, since it changes the deadline for every student in the section. A starting tool set: list_assignments, extend_deadline, list_submissions (with late and missing filters), message_students.

Travel booking

"flights to Lisbon under €200 next weekend"

{ "name": "search_flights", "arguments": { "destination": "Lisbon", "max_price": 200, "currency": "EUR", "depart": "next weekend" } }

A filter. A starting tool set: search_flights, search_hotels, list_bookings, cancel_booking. Dates such as "next weekend" can stay as text for your code to resolve, or you can train the model on exact dates.

Marketing automation

"pause the spring sale emails for contacts in Germany"

{ "name": "pause_campaign", "arguments": { "campaign": "spring sale", "country": "DE" } }

An action. A starting tool set: list_campaigns, pause_campaign, resume_campaign, list_contacts (with country, segment and activity filters). Country names map well to an enum of codes.

Smart home and devices

"dim the living room lights to 30%"

{ "name": "set_light", "arguments": { "room": "living room", "brightness": 30 } }

An action. A starting tool set: set_light, set_thermostat, lock_door, run_scene. Calls like these are short and the tool set is small, which suits Fern Bud (12M), the smallest model.

What makes a domain a good fit

  • The requests repeat. People ask for the same kinds of thing in different words.
  • The actions already exist. Your API has the operations, and the model only has to choose one and fill its arguments.
  • Each request is one step. "Find these and change them" is one call. "Research this and write a summary" is not.
  • A wrong call is visible. A user can see and fix a call before it runs, or the action can be undone.

Training best practices covers how to describe the tools so the model learns them well.

Revised

Was this page helpful?