Document extraction

Apart from POST /v1/systemone, you can use the extraction API, POST /v1/extract, which is built to extract structured data from PDFs and images. You don’t write the questions yourself: you describe the fields you want, and the API finds each one in the document and returns it in the type you asked for.

Extracting fields from a document

Send a PDF or an image together with a JSON Schema that lists the fields you want. For each field, Strom finds the line in the document that contains it. Scanned pages and images are converted to text using OCR automatically.

curl https://api.uprelic.com/v1/extract \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "strom-1.0.7",
    "schema": {
      "properties": {
        "invoice_number": {"type": "string"},
        "invoice_date": {"type": "string", "format": "date"},
        "total": {"type": "number", "format": "amount", "description": "The total including tax"},
        "paid": {"type": "boolean"}
      }
    },
    "document": {"url": "data:application/pdf;base64,JVBERi0xLjQK..."}
  }'

document.url can be a public URL or a base64 data: URL. PDFs can have at most 8 pages. model can be strom-fast or strom-base (beta), or their versioned names; strom-base is much slower and costs more.

Each field can have a description. Use it to tell the model what to look for when the name alone is ambiguous, for example whether total includes tax.

API Response

For each field, the response contains:

• text: the line of the document the field was found on, exactly as it appears.
• value: that line converted to the field’s type, for example a date as 1988-12-16 or an amount as 14982.0. See the table below.
• confidence: how sure the model is that it picked the right line, from 0 to 1.
• page and box: where the line is. box is [left, top, right, bottom] as fractions of the page’s width and height.

If the document doesn’t contain a field, that field is null.

{
  "model": "strom-1.0.7",
  "fields": {
    "invoice_number": {
      "text": "17795", "value": "17795", "confidence": 0.98,
      "page": 1, "box": [0.7731, 0.1286, 0.8225, 0.1424]
    },
    "invoice_date": {
      "text": "12/16/88", "value": "1988-12-16", "confidence": 0.94,
      "page": 1, "box": [0.7731, 0.1552, 0.85, 0.1705]
    },
    "total": {
      "text": "$14,982.00", "value": 14982.0, "confidence": 0.55,
      "page": 1, "box": [0.8306, 0.3529, 0.9256, 0.3662]
    },
    "paid": {"text": null, "value": false, "confidence": 0.58, "page": null, "box": null}
  },
  "usage": {
    "input_tokens": 20060, "output_tokens": 5, "cached_input_tokens": 25344,
    "cost": 0.00084252, "currency": "usd"
  },
  "metadata": {"server_processing_time_ms": 3405}
}

Field types

The field’s type in the schema decides how text is converted into value:

If the line contains more than one date, number or percentage, we don’t guess which one you meant: value is then the whole line, the same as text.

Cost and errors

• You pay for input tokens at the same price as POST /v1/systemone. The cost of each extraction is in usage.cost. For example, a 2-page PDF invoice with a 10-field schema used about 28,000 input tokens.
• The “Improve document extraction” switch under Billing comes with the paid plans. Each run that improves your extraction is charged to your balance for the AI processing it uses, at twice our provider's price, about €2 to €3 a run.
• Documents that can’t be read are not charged. You get a 422 with the reason in detail, for example when the file isn’t a PDF or image, has no text, or has more than 8 pages.