Ctrlk
Tonic.ai
For the complete documentation index, see llms.txt. This page is also available as Markdown.

Redact individual strings

Required global permission: Use the API to parse or redact a text string

Before you perform these tasks, remember to instantiate the SDK client.

You can use the Tonic Textual SDK to redact individual strings, including:

  • Plain text strings

  • JSON content

  • XML content

For a text string, you can also request synthesized values from a large language model (LLM).

The redaction request can include the handling configuration for entity types.

The redaction response includes the redacted or synthesized content and details about the detected entity values.

Redact a plain text string

To send a plain text string for redaction, use textual.redact:

redaction_response = textual.redact("""<text of the string>""")
redaction_response.describe()

For example:

redaction_response = textual.redact("""Contact Tonic AI with questions""")
redaction_response.describe()

Contact ORGANIZATION_EPfC7XZUZ with questions
    
{"start": 8, "end": 16, "new_start": 8, "new_end": 30, "label": "ORGANIZATION", "text": "Tonic AI", "new_text": "[ORGANIZATION]", "score": 0.85, "language": "en"}

The redact call provides an option to record the request, to allow you to preview the results in the Textual application. For more information, go to .

Redact multiple plain text strings

To send multiple plain text strings for redaction, use textual.redact_bulk:

For example:

Redact JSON content

To send a JSON string for redaction, use textual.redact_json. You can send the JSON content as a JSON string or a Python dictionary.

redact_json ensures that only the values are redacted. It ignores the keys.

Basic JSON redaction example

Here is a basic example of a JSON redaction request:

It produces the following JSON output:

Specifying entity types for specific JSON paths

When you redact a JSON string, you can optionally assign specific entity types to selected JSON paths.

To do this, you include the jsonpath_allow_lists parameter. Each entry consists of an entity type and a list of JSON paths for which to always use that entity type. Each JSON path must point to a simple string or numeric value.

The specified entity type overrides both the detected entity type and any added or excluded values.

In the following example, the value of the key1 node is always treated as a telephone number:

It produces the following redacted output:

Redact XML content

To send an XML string for redaction, use textual.redact_xml.

redact_xml ensures that only the values are redacted. It ignores the XML markup.

For example:

Produces the following XML output:

Redact HTML content

To send an HTML string for redaction, use textual.redact_html.

redact_html ensures that only the values are redacted. It ignores the HTML markup.

For example:

Produces the following HTML output:

Synthesis with Large Language Models

You can request synthesized values from a large language model (LLM) using two different approaches.

ReplacementSynthesis

ReplacementSynthesis redacts sensitive values and uses the LLM to generate realistic replacements based on the surrounding context.

When you use this process, Textual first identifies the sensitive values in the text. It then sends the value locations and redacted values to the LLM. For example, if Textual identifies a product name, it sends the location and the redacted value PRODUCT to the LLM. Textual does not send the original values to the LLM.

The LLM then generates realistic synthesized values of the appropriate value types based on the context of the surrounding text.

Example:

Output:

GroupingSynthesis

GroupingSynthesis groups related entities, generates new entity names, then uses the LLM to reproduce the original format of the value.

This approach is particularly useful when values have specific formats that need to be preserved. For example, if a name is spelled out using the phonetic alphabet (e.g., "B as in boy, O as in orange, B as in boy" for "Bob"), GroupingSynthesis will:

  1. Identify the grouped entity ("Bob")

  2. Generate a new entity name without using LLM ("Tom")

  3. Use the LLM to reproduce the same format ("T as in toy, O as in orange, M as in mark")

Example:

Output:

Configuration

Use the generator_config parameter to specify which entity types should use synthesis and which synthesis method to apply. Use generator_default to set the default behavior for entity types not explicitly configured.

For more information about configuring entity type handling, see Configure entity type handling for redaction.

Note: Before you can use either synthesis method, you must enable additional LLM processing. The additional processing sends the values and surrounding text to the LLM. For an overview of the LLM processing and how to enable it, see the documentation about configuring the Solar.LLM container inConfiguring the Solar.LLM container

Format of the redaction and synthesis response

The response provides the redacted or synthesized version of the string, and the list of detected entity values.

For each redacted item, the response includes:

  • The location of the value in the original text (start and end)

  • The location of the value in the redacted version of the string (new_start and new_end)

  • The entity type (label)

  • The original value (text)

  • The replacement value (new_text). new_text is null in the following cases:

    • The entity type is ignored

  • A score to indicate confidence in the detection and redaction (score)

  • The detected language for the value (language)

  • For responses from textual.redact_json, the JSON path to the entity in the original document (json_path)

  • For responses from textual.redact_xml, the XPath to the entity in the original XML document (xml_path)

Last updated

Was this helpful?