Datasets workflow for text redaction and synthesis

You can use Textual to generate versions of files where sensitive values are synthesized or redacted.

To only generate redacted files, you use a Tonic Textual dataset.

You can also optionally configure a Textual pipeline to generate redacted files in addition to the JSON output.

At a high level, to use Textual to create synthesized or redacted data:

Diagram of the Tonic Textual redaction and synthesis workflow

1. Create a Textual dataset or pipeline. A dataset is a set of files to redact. A pipeline is used to generate JSON output that can be used to populate an LLM system. Pipelines also provide an option to generate redacted versions of the selected files.

2. Add files to the dataset or pipeline. Textual supports almost any free-text file, PDF files, .docx files, and .xlsx files. For images, Textual supports PNG, JPG (both .jpg and .jpeg), and TIF (both .tif and .tiff) files.

3. For a dataset or an uploaded files pipeline, as you add the files, Textual automatically uses its built-in models to identify entities in the files and generate the pipeline output. For an Amazon S3 or Databricks pipeline, to identify the entities and generate the output, you run the pipeline.

4. For a dataset, review the types of entities that were detected across all of the files. For pipeline files, the file details include the entities that were detected in that file.

5. Configure how to handle each type of value. By default, Textual redacts the entity values, which means to replace the values with a placeholder that identifies the type of sensitive value. For example, PERSON, LOCATION. For PDF files and image files, redaction means to cover the value with a black box. For a given entity type, you can instead choose to synthesize the values, which means to replace the original value with a realistic replacement. You can also choose to ignore the values, and not replace them. For a dataset, Textual automatically updates the file previews and downloadable files to reflect the updated configuration.

   For a pipeline, the updated configuration is applied the next time you run the pipeline, and only applies to new files.

6. Optionally, in a dataset, you can create a list of values to not identify as a specific entity type, if some of the detected values are incorrect. Pipelines do not allow you to exclude individual values.

7. Datasets also provide additional options for PDF files. These options are not available in pipelines. You can add manual overrides to a PDF file. When you add a manual override, you draw a box to identify the affected portion of the file.

   You can use manual overrides either to ignore the automatically detected redactions in the selected area, or to redact the selected area. To make it easier to process multiple files that have a similar format, such as a form, you can create templates that you can apply to PDF files in the dataset.

8. You then download the redacted and synthesized versions of the files to use as needed.