What is Fernfly
Fernfly trains a small model on your API, so a sentence a user types becomes a call your app can run.
Fernfly turns a sentence into an action in your app. A user types "unpaid invoices over $500 this quarter", and a model trained on your API returns the call that answers it:
{
"name": "list_invoices",
"arguments": { "status": "unpaid", "min_total": 500, "issued": "this_quarter" }
}Your app runs that call with its own code and the signed-in user's permissions. Fernfly only decides which call to make. It never runs one.
The three products
| Product | What it is |
|---|---|
| Intent bar | A bar your users type into. It turns what they type into filters, bulk edits and actions in your app, and shows what it will do before it does it. |
| Fernfly Studio | Where you build and run a model: read your API docs, write training examples, train, deploy, and improve the model from real traffic. |
| Fern models | The small models Studio trains: Fern Bud (12M), Fern Pinnule (160M) and Fern Pinna (1.5B). |
You can use Studio and a Fern model without the intent bar. Your own backend can send sentences to the model's endpoint and act on the calls it returns.
How it works
Describe your API. Give Studio a link to your API docs or an OpenAPI file. It lists your API's operations as tools.
Write the training data. Studio writes example sentences for each tool, each paired with the call it should produce. You review them.
Train. Studio fine-tunes a Fern model on those examples. This takes minutes for Fern Bud and longer for the larger models.
Deploy. Studio puts the trained model behind an endpoint as soon as training finishes.
Use it. The intent bar, or your own code, sends each sentence to the endpoint and runs the call that comes back.
Improve it. Requests the model missed show up in Studio. You correct them and train again.
Terms used in these docs
| Term | Meaning |
|---|---|
| Tool | One operation your app can perform, such as list_invoices, described by a name, a description and its arguments. |
| Tool call | The model's answer: a tool's name and the values of its arguments. |
| Utterance | The sentence a user typed. |
| Training pair | One example the model learns from: an utterance and the tool call it should produce. |
| Fine-tuning | Training an existing model further on your own examples, so it learns your tools. |
| Base model | The Fern model a fine-tune starts from. |
| Deployment | A trained model put behind an endpoint so it can answer requests. |
| Endpoint | The URL your app sends utterances to. |
| AI fallback | A second attempt by a general-purpose model when your model can't produce a tool call. |
Where to start
Revised