Playground

The Playground in the dashboard runs extractions without writing any code. Drop in text or images, define a schema or a prompt, pick an engine, and see the structured output plus its exact cost. It is the fastest way to design and tune a schema before wiring the API into your app.

Playground runs are real jobs — they appear in your job history with source: "playground" and are billed the same way as API jobs.

Two modes

• Schema mode — provide a JSON Schema and get strict, typed JSON back. Best when you know the fields you want.
• Prompt mode — give a natural-language global_prompt and a target of markdown, plaintext, or schemaless json. Best for summaries, freeform structuring, or exploration.

Whatever you build in the Playground maps one-to-one to the [POST /jobs/create](/docs/api-reference) body, so you can copy a working setup straight into code.

Writing schemas that steer the model

Veralens uses JSON Schema directly — the schema both shapes the output and steers the model. A few rules make extractions dramatically more reliable:

• Describe every field. description is the primary steering lever. State what the field means, its units, and the exact format you want in plain language (e.g. *"Issue date in ISO format YYYY-MM-DD, date only"*).
• **Use enum for classification.** For a document type, status, or any closed set, list the allowed values and the model must pick one.
• **Describe formats — don't rely on pattern.** Validation keywords like pattern, minLength, maxLength, minimum, and multipleOf are stripped before the schema reaches the model. Put format rules in description instead (e.g. *"Exactly 5 digits"*, *"Phone with country code, e.g. +1 555 123 4567"*).
• **type, properties, items, required, enum, format** are respected and define the structure.

{
  "type": "object",
  "properties": {
    "documentType": {
      "type": "string",
      "description": "Classify the document into exactly one of these types.",
      "enum": ["invoice", "receipt", "contract", "form", "letter", "other"]
    },
    "issueDate": {
      "type": "string",
      "description": "Issue date in ISO format YYYY-MM-DD. Date only, no time."
    },
    "lineItems": {
      "type": "array",
      "description": "Each row on the document.",
      "items": {
        "type": "object",
        "properties": {
          "label": { "type": "string", "description": "Item description as printed." },
          "amount": { "type": "number", "description": "Line amount. Dot as decimal separator." }
        }
      }
    }
  },
  "required": ["documentType"]
}

Dates, currencies, and country codes in your output are normalized to ISO 8601, ISO 4217, and ISO 3166 automatically, so you rarely need post-processing for those.

From Playground to production

1. Iterate on the schema (or prompt) until the output is right.
2. Note the engine you used — vera-1.0-high for accuracy, vera-1.0-low for cheaper high-volume runs.
3. Copy the same sources / schema / target shape into a [POST /jobs/create](/docs/api-reference) call.
4. Add a [webhook_url](/docs/jobs) or a polling loop to collect results.