Docs

Create a model

Start a project from your API docs, an OpenAPI file, a list of tools or a template, and take it through to a deployed model.

The new-project wizard takes a project from a description of your API to a deployed model. Open it with New project in the sidebar.

Choose how to start

The wizard asks How do you want to start? and offers three options.

OptionUse it whenWhat happens
Intent bar for your appYou want the intent bar in your app and have API docs or an OpenAPI file.Studio runs every step on its own, trains on Fern Pinna (1.5B) and ends on the intent bar setup.
From scratchYou want to control each step, or to use the model from your own code.You go through the steps one at a time.
WooCommerceYou run a WooCommerce store.Studio fills in 28 store tools and over 5,500 training pairs, and takes you to the Train step.

The wizard has five steps: Source, Extract, Generate, Train and Deploy. Both the intent bar start and From scratch go through all five.

The intent bar start

Choose Intent bar for your app.

Under Your API docs, choose Docs link and paste the address of your API reference, or choose OpenAPI file and upload a .json or .yaml spec. The docs must be public: Studio can't read a page that needs signing in.

Optionally, enter Your app's address. Only that site may then call your model without a key, and Studio reads its public pages to add them to the intent bar's instant search. You can change this later.

Optionally, enter a Project name. Left blank, the project is named after your docs.

Choose Build my intent bar.

Studio then reads the docs, writes training pairs, trains and deploys, one after another. A bar above the steps shows which step is running and the time so far. Keep the tab open: the model is deployed from this page when training ends.

The run stops for your input in one case. If your docs describe more than 30 operations, Studio selects the 30 your users are most likely to ask for and waits. Change the selection if you like, then choose Continue building. An intent bar works best with at most 30 tools, and the wizard won't continue with more.

If a step fails, the run stops there and that step's own controls take over, so you can fix the problem and carry on from the same step.

When the run finishes, the Deploy step becomes Add the intent bar. Continue with Build an intent bar.

From scratch

1. Source

Choose From scratch, then a Source type:

Source typeWhat to give it
OpenAPI / SwaggerUpload a spec (.json, .yaml or .yml).
Route ManifestA JSON list of your tools, pasted on the next step.
API docs URLThe address of a public docs page or an OpenAPI file.

GitHub Repo is marked Coming soon. Name the project if you like, then choose Continue.

2. Extract

Choose Extract routes. Studio reads your source and lists every operation it found as a tool.

For an API docs URL, Studio finds the pages of your reference, reads them, and lists each operation. It only keeps an operation when the endpoint appears in your docs, so it doesn't invent endpoints. A progress list shows each stage. A link to an OpenAPI or Swagger file, or a reference page built on one (Redoc, Swagger UI, RapiDoc or Stoplight), is read as a spec instead.

For a route manifest, paste a JSON array into Tools JSON. Each tool needs a name, a description and, if it takes arguments, a parameters object in JSON Schema form. Show expected format has an example:

[
  {
    "name": "search_products",
    "description": "Search the product catalog",
    "parameters": {
      "type": "object",
      "properties": {
        "query": { "type": "string", "description": "What to search for" },
        "limit": { "type": "number", "description": "Max results to return" }
      },
      "required": ["query"]
    }
  }
]

Review the list. Untick anything your users won't ask for, such as admin or internal operations. Select recommended picks the 30 Studio considers most useful when there are more than that. Then choose Continue.

3. Generate

Studio writes training pairs for the selected tools with Gemini. First choose an Utterance style:

StyleHow the examples are writtenBest for
VerboseFull, free-flowing sentences.Users who type whole sentences or describe what they want.
TerseShort commands built from templates, with the arguments filled in by code so they are always exact.A command bar with short input. Usually more accurate at picking the tool.

Then choose how many pairs to write per tool:

  • Verbose: 10, 15, 25 (recommended) or 50 per tool. More pairs teach the model more ways of saying the same thing, and take longer to write.
  • Terse: Quick test, Baseline, Recommended or Maximum. Terse fills its templates without further Gemini calls, so it can produce many more pairs within your daily limit.

Choose Generate with Gemini. Studio also adds examples of requests your tools can't answer, so the model learns to say so instead of guessing. When it finishes, the pairs appear in a table of Input and Output.

To use pairs you already have, choose Import JSONL instead. See Training data for the format. Export JSONL downloads the pairs.

Choose Continue when you're happy with the pairs.

4. Train

Choose the hardware. A100 is included in every plan. H100 is marked Coming soon.

Advanced mode shows the settings a run starts from. It is closed by default, and the fields are already filled in for your project, so you can leave it alone:

SettingWhat it is
Base modelThe Fern model to fine-tune: Fern Bud, Fern Pinnule or Fern Pinna. See Fern models.
EpochsHow many times training passes over all your pairs.
Batch sizeHow many pairs the model learns from in each update.
Learning rateHow large each update is.

These settings are fixed by the project's first run. Retraining reuses them.

Choose Start training. The Training log shows each step, and a Loss curve appears as training goes on: the line falls as the model learns. Stop ends the run early and keeps the model trained so far. You then choose Keep this model to use it, or Discard to start again.

5. Deploy

Studio deploys the model as soon as training finishes. The Deploy step shows:

  • Endpoint, the URL to send utterances to,
  • API key, shown once. Save it now: Studio can't show it again,
  • Keyless access, which lets a browser call the model without a key. See API keys and keyless calls.

Choose Open project to go to the project's Overview, where Try it sends your first request.

Starting from the WooCommerce template

Choose WooCommerce, then Set up template & continue. Studio adds the template's tools and training pairs and takes you to the Train step, with the base model and settings already tuned for the template. You can step back to review the tools and pairs before you train.

Leaving and coming back

You can close the wizard after any step. The project keeps its progress, and its Overview shows Continue setup until it has training pairs. A training run carries on without the wizard. Follow it on the project's Training jobs tab.

Revised

Was this page helpful?