AI: Query Markdown content

The "AI: Query Markdown content" service task answers targeted questions about a Markdown text and returns the answers as a structured JSON object. Questions are passed as an array, with each question linked to an attribute name — automatically producing a clean result object.

This task is particularly well suited for analyzing OCR results: a scanned PDF is first converted to Markdown, then specific information such as delivery note numbers, dates or amounts can be extracted through targeted questions.

---

Input parameters

Provide the following fields as task input:

{
  "content": "# Delivery Note\n\nDelivery Note No: DN-2024-001\nCommission: COM-4711\nDate: 15 March 2024\n\n| Pos | Article | Qty |\n|-----|---------|-----|\n| 1 | Widget A | 100 |\n| 2 | Widget B | 50 |",
  "questions": [
    {
      "question": "What is the delivery note number?",
      "attribute": "deliveryNoteNumber",
      "format": "e.g. DN-2024-001"
    },
    {
      "question": "What is the commission number?",
      "attribute": "commissionNumber"
    },
    {
      "question": "What is the date of the delivery note?",
      "attribute": "date",
      "format": "YYYY-MM-DD",
      "validation": "^\\d{4}-\\d{2}-\\d{2}$"
    },
    {
      "question": "How many line items does the table contain?",
      "attribute": "itemCount",
      "format": "Integer"
    }
  ],
  "returnNotFoundFields": false,
  "maxContentLength": 50000
}

Explanation:

• content: The Markdown text to be analyzed. Typically an OCR result or other structured text from a previous process step.
• questions: An array of questions. Each question contains:
  • question: The question in natural language (English or German).
  • attribute: The key name in the result object.
  • format (optional): A hint for the AI about the expected answer format (e.g. "YYYY-MM-DD" or "integer").
  • validation (optional): A regular expression (regex) to validate the extracted value. If the value does not match, the AI is automatically asked to correct it (up to 3 attempts).
• returnNotFoundFields (optional, default: false): If true, the result includes an additional _notFound field listing attribute names for which no answer was found.
• maxContentLength (optional, default: 50000): Maximum character count of the content. Can be increased if needed.

---

Output

The task returns a JSON object whose keys match the attribute names of the questions:

{
  "deliveryNoteNumber": "DN-2024-001",
  "commissionNumber": "COM-4711",
  "date": "2024-03-15",
  "itemCount": 2
}

If an answer cannot be found in the text, the field is set to null.

With returnNotFoundFields: true:

{
  "deliveryNoteNumber": "DN-2024-001",
  "commissionNumber": "COM-4711",
  "date": "2024-03-15",
  "itemCount": 2,
  "_notFound": []
}

---

JSONata examples

{
  "content": ocrResult.markdown,
  "questions": [
    {
      "question": "What is the delivery note number?",
      "attribute": "deliveryNoteNumber",
      "format": "e.g. DN-2024-001"
    },
    {
      "question": "What is the date?",
      "attribute": "date",
      "format": "YYYY-MM-DD",
      "validation": "^\\d{4}-\\d{2}-\\d{2}$"
    }
  ]
}

{
  "content": previous_step.text,
  "questions": [
    {
      "question": "What is the total amount?",
      "attribute": "totalAmount",
      "format": "Number"
    },
    {
      "question": "What is the invoice number?",
      "attribute": "invoiceNumber"
    }
  ],
  "returnNotFoundFields": true
}

---

Notes

• At least one question must be included in the questions array.
• Questions can be written in any language — the AI understands both English and German equally well.
• Precise format hints significantly improve result quality (e.g. "Date in YYYY-MM-DD format" rather than just "date").
• The optional validation enables automatic verification: if the extracted value does not match the regex pattern, the AI is re-queried with targeted correction feedback.
• For very long texts, maxContentLength can be increased. The default of 50,000 characters covers most use cases.

---

Tip

This task works particularly well in combination with OCR results: first convert a scanned PDF to Markdown via OCR, then use this task to extract the relevant information. Combined with the "AI: Extract key-value pairs" task, a complete document intake workflow can be built — automatic detection of all fields plus targeted follow-up queries for critical values.