Import from Unstructured Data

The Import from Unstructured Data task allows importing data directly inside a flow, by extracting a tabular dataset out of a single unstructured file, such as a PDF, a Word document, an HTML page or any other generic text-based file, using a Large Language Model.

Once opened, the task interface has a single Options tab, divided into two areas:

Warning

Always remember to save and compute the task.


The Options tab

The Options tab follows the structure described in the Natural Language tasks family page, for the Model controller and its common connection parameters.

Source file

Once the AI provider connection has been configured, the file to be imported must be selected.

The connection to the filesystem containing the file, and the file selection itself, follow the same Saved/ Custom logic used for the AI Model connection and are described in the Filesystem resource page. Unlike other import tasks, only a single file can be selected at a time, and no restriction is applied on its extension: PDF, Word, HTML pages, JSON files and any other generic text-based file are all accepted.

See also

For further information about configuring a filesystem connection, refer to the Filesystem resource page.

Prompt

  • Prompt: the instructions given to the model, describing how the unstructured content of the selected file has to be interpreted and converted into a tabular dataset, e.g. describing the columns to be extracted and their meaning.


Outputs

Once the task has finished running, it produces a dataset output, containing the tabular data extracted from the selected file, according to the instructions provided in the Prompt.


Example

  • Configure the Options tab by choosing an AI provider and filling in its connection parameters.

  • Select the source file to import, for example a PDF file containing a table of product prices.

  • Write a prompt describing how the content of the file should be converted into a dataset, for example: “Extract from this document a dataset with one row per product, with columns Product, Price and Currency.”

  • Save and compute the task.

Once computed, the task returns a dataset with one row per product, and the columns Product, Price and Currency, populated with the information extracted from the unstructured content of the source file.