Some features in this tutorial require **Commercial** or **Enterprise** licenses.

This tutorial assumes you've followed the tutorials described in the **Create and share an app** sequence of tutorials, up through **Defining fields.**

If you're struggling to extract a specific piece of information, there are a few different levers you can use to improve results. This guide describes the available options.

Click **Add field** to see all the options for fine-tuning extraction

For a simple extraction field, add an optional **description**

To do this, click the **Description** field and add a natural language description of the information you want to extract, the description will be passed along with the field name to the AI model

This additional information helps the model locate the information you're looking for

If the format of the extracted information isn't correct, you can use an optional **cleaning function** to convert the data into your preferred format

Cleaning functions are also useful for whittling down large pieces of extracted text, you can reduce a paragraph, table, or document section down to the specific piece of information you're looking for

To add a cleaning function, click **Add cleaning** and pick one of the cleaning options

AI Hub provides a number of different pre-built cleaning options, you can also create a custom cleaning prompt or function

Click **Add cleaning** again to hide the cleaning options

There are two AI models available. Switching to the advanced model might improve results, but is slower and costs more consumption units per extraction

Click **Change model** to switch between AI models

Click **Change model** again to hide the model options

If you need more extraction power or flexibility, you can define the field with a **Reasoning prompt**, this lets you use natural language to specify a task for the model to perform in order to extract the right data

This is particularly useful if you want to extract long-form data, or if the model needs to perform additional problem solving or reasoning in order to get the right information

Let's make a field defined by a reasoning prompt next

Used to generate values that must be deduced based on information in the document.

Click the **Natural language prompt** field

If the prompt already contains default text, erase it

Type [[Are there transactions related to gambling or betting? Return a yes or no answer.]]

Note that you're not only telling the app what *information* should be extracted into the field, but also what *format* it should be in

The field's result is [[Yes]] because there the AI model detected transactions related to gambling or betting

Close the **Gambling Detected** field definition by clicking the **X**

The final option for improving results is to change the data type of a field

Wherever you see the **Table** symbol in a document, it means the AI model has detected that a table might be present

We can extract these tables by defining a field with the **Table extraction** data type

Make a new field by clicking **Add field**

Click the **Field Type** dropdown and change its value from **Text extraction** to **Table extraction**

In a few seconds the project will show a snippet of the extracted table in the field's **Result** box

Click the **Maximize** button to see the whole table

Here's the entire extracted table of transactions

The other data type available for extraction is a **List extraction**

This is useful when there are multiple pieces of information you want to extract, and they don't already exist in a table within the original document

To change this field to use the **List extraction** data type, click the **Field type** dropdown

To configure a **List extraction** field, enter a field description and optional attributes

Attributes are the headers in the output table that you want the field to generate, It's a good idea to experiment with them to get a better feel for how they work

Instabase

This tutorial describes options for improving extraction results. It covers field descriptions, cleaning functions, AI model options, reasoning prompts, and data types.

Improve results