Prompts

A prompt is the instruction sent to the model for each document. This guide shows how to design prompts that produce clean, structured output, and how to get multiple rows of output per document using long format.

Structured outputs

Use structured JSON output when you want to extract specific fields from your documents for further processing. Make sure your prompt also tells the model how to format the output.

In this example, we extract the following fields: title, authors, published date, category, page count, and whether the document contains pictures. Here’s the prompt:

Please extract the following information from the provided document. Do not use any other sources.

The document title. All authors separated by a comma — if there are no authors, set the field to null. The date the document was
published in the format "DD.MM.YYYY" — if unknown, set the field to null. The number of pages. Whether the document includes any pictures.

Example JSON output:
{
    "title": "A document title",
    "authors": "Ben, Lisa",
    "published_date": "10.12.2004",
    "page_numbers": 14,
    "pictures_included": true
}

It’s important to tell the model exactly how to return the data. When using JSON output, including the word json in your prompt is often required.

Export the data

Once your batch is done, the model’s consistent JSON output can be converted into a spreadsheet. Click the Export Icon and select Convert JSON CSV (or Convert JSON Excel) as the export option. Both convert your JSON output into columns, one per field.

Here’s what the table looks like for this example:

file_name	_title	_authors	_published_date	_page_numbers	_pictures_included
example-1.pdf	A document title	Ben, Lisa	10.12.2004	14	True
example-2.pdf	Another document title	null	null	1	False
example-3.pdf	Beautiful title	Fine	17.02.2016	7	True

Why the underscores _ before column names?

Underscores are added automatically to avoid naming conflicts with system columns.

For example, if your output had a field called id, it would clash with the system id column. To keep things clear: columns without an underscore contain system information (like model name and temperature), while columns with an underscore come from your structured output.

Long format

By default, each document produces one row of output. Sometimes you want multiple rows per document instead — for example, one row per question you asked. This is called long format. It’s useful in scientific evaluation when you need to compare different data sets.

There are two ways to get long format output:

A) Multi-row interpretations

Instruct the model itself to return one answer per row, all in a single request.

Say you have three questions:

1. Is this a scientific document?
2. Does this document provide sources?
3. Was this document published before 2020?

Target output

file_name	_id	_answer
example-1.pdf	1	True
example-1.pdf	2	False
example-1.pdf	3	False
example-2.pdf	1	False
example-2.pdf	2	True
example-2.pdf	3	False

Prompt instructions

To put each answer in its own row, instruct the model to return a JSON array with one element per question.

Here’s an example prompt:

Please answer the following questions.
1. Is this a scientific document?
2. Does this document provide sources?
3. Was this document published before 2020?

Return your answers in this JSON format:
{
    [
        { "id": 1, "answer": True  },
        { "id": 2, "answer": False },
        { "id": 3, "answer": True  }
    ]
}

B) Prompts with multiple steps

Instead of asking the model to produce several rows in one response, split your prompt into several steps (complex prompts), each sent as its own request. Each step gets its own row.

Complex prompts let you split one prompt into several steps, each with its own context. This is useful for complex tasks or smaller models that struggle with multiple questions at once.

To get started, click + Add Prompt and select Complex instead of Simple. You can then choose how you want to build it:

• Visual Editor — fill in the pre-prompt, post-prompt, and each step through dedicated fields.
• YAML — write the configuration yourself. Use the following format:

multi_prompt_v1:
  pre: "You are a helpful assistant."
  post: "Please only refer to the provided document."
  tasks:
    - id: step_1
      prompt: |
        What is the document title?
    - id: step_2
      prompt: |
        Does the document have any authors?

Both options produce the same result — pick whichever is more convenient. Instead of one request per document, two API requests will be made. For this example, the prompts will look like this:

First prompt

You are a helpful assistant.
What is the document title?
Please only refer to the provided document.

Second prompt

You are a helpful assistant.
Does the document have any authors?
Please only refer to the provided document.

Caution

Complex prompts increase your costs. In this example, costs double because two requests are made per document.

What is the ID for?

The ID helps you identify which prompt produced which result. It appears in the prompt_marker column when you export your data.