Train Classifier

jina.train_classifier
ActionAI & modelsconnection required

Calls the Jina AI API with the connection's credentials.

Train a few-shot classifier from labelled examples, or add more examples to a classifier that already exists.

Connection

Required
Yes — the node will not save without one
Auth type
Bearer token
Providers
Jina AI
Secret shape
the API key, or {"api_key": "…"}
Config it needs
base_url, default_embedding_model, default_reranker_model
Resolved by
config.connection_id, else node.connection_id

Configuration

5 fields, 1 required. A dotted path is a nested object and name[] an array of objects — the editor renders the same tree.

FieldTypeRequiredDefaultConstraintsExample
input
Labelled examples, each {"text": "…", "label": "…"}. Creating a classifier needs at least two distinct labels.
arrayYes—array of object[{"label":"billing","text":"refund me"}]
model
Embedding model to build a new classifier on, for example jina-embeddings-v3. Ignored when classifier_id is set.
stringNo——jina-embeddings-v3
classifier_id
Add these examples to an existing classifier instead of creating one.
stringNo———
access
Visibility of a newly created classifier.
stringNopublicone of "public", "private"—
num_iters
Training iterations to run over the examples.
numberNo101–100—
  • At least one of model, classifier_id must be set.

Example configuration

Every value below is the contract’s own example or default — nothing invented.

{
  "access": "public",
  "input": [
    {
      "label": "billing",
      "text": "refund me"
    }
  ],
  "model": "jina-embeddings-v3",
  "num_iters": 10
}

Output

One shape.

result

FieldTypeExample
classifier_id
The classifier that was trained.
string—
num_samples
Examples accepted in this request.
number—
created
True when this call created the classifier.
boolean—
usage
Token usage Jina billed for this call.
object—
{
  "classifier_id": "",
  "created": false,
  "num_samples": 0,
  "usage": {}
}

See Train Classifier run in a workflow.

In a technical demo we add the Jina AI connection, drop the node on the canvas and let Workflow Health pre-flight it before the first real run.