Redactwizard – Help

Version 2.5

Redactwizard is a desktop application for detecting, redacting, anonymizing, and pseudonymizing documents. PDF, text, and supported document-processing tasks are handled locally on your computer. Integrated local AI models help identify potentially sensitive information.

Redactwizard is an assistance tool, not a guarantee that every sensitive detail will be detected. Automated detection can miss relevant information or produce false positives. Always review the complete output before sharing or publishing a document.

1. Getting Started: Portable Version

Redactwizard is provided as a portable application. A conventional installation is not required.

Start Redactwizard directly from the portable application folder. The application keeps its required runtime and configuration files alongside the application in its working directory. Avoid separating the executable from its associated files.

Recommendation

Keep the complete Redactwizard folder together and make backups of any configuration or data you need to preserve. Do not move or copy individual components separately.

Windows Security Notice: Microsoft Defender SmartScreen

Windows may display a Microsoft Defender SmartScreen warning when Redactwizard is started for the first time, for example because the application is new or has not yet built up a reputation with Microsoft.

Step 1: If the Windows security screen appears, select More info.

Review the publisher and file information. Proceed only if the application came from the source you intended to use and you have verified its origin.

2. Missing Microsoft Visual C++ Runtime

In some environments, Windows may report that a Microsoft Visual C++ Runtime component is missing. This can prevent the application or one of its native libraries from starting correctly.

Use only the official Microsoft package for Windows x64:

Download Microsoft Visual C++ Redistributable x64

After installation, restart Redactwizard. Restart Windows as well if required.

Security recommendation

Do not download individual DLL files or runtime components from third-party download portals.

3. Start Screen and Basic Layout

After startup, the central workspace is displayed. When no document is open, this area acts as a dashboard for starting a task.

The top bar provides access to application functions and document opening. Available tools depend on the active product edition and document type.

• Redactwizard Legal 2: Provides PDF and text processing, batch processing, and shared management and settings functions.
• Redactwizard Pro 2: Adds the Office Editor, AI Chat, native vector redaction, and additional local AI models.

Documents remain open in their tabs when you switch between workspaces. This allows you to work on several documents without reopening them.

Recommendation

Use separate workspaces for unrelated matters or projects to help keep documents from different clients or cases organized.

4. Workspaces, Desktops, and Tabs

Redactwizard provides four virtual workspaces, each with its own tab bar.

Use the plus icon to open another editor or work item. Open documents appear as tabs; select a tab to activate the corresponding document.

Switch between workspaces using the controls in the upper-right corner. Tabs remain open in their respective workspaces. You can also switch workspaces using Ctrl+1 through Ctrl+4.

Recommendation

For example, keep one matter in the first workspace, another matter in the second, and documents awaiting review in the third.

5. Opening Files

You can open files from the application menu, using the folder button in the tab bar, or through the open function on the start screen.

The file extension determines which editor is used.

The Office Editor is available only in Redactwizard Pro 2. Redactwizard can also retrieve documents through the Clio integration when configured or acquire a document using the integrated scanner.

Choosing the appropriate editor

Use the PDF Editor for existing PDFs and the Text Editor for plain-text files. Use the Office Editor for supported Office and OpenDocument files. If a file has unusually complex formatting, consider a PDF workflow only when conversion is appropriate and the resulting PDF can be checked carefully.

PDF Editor

6. PDF Editor – Overview and Basic Workflow

The PDF Editor is the main workspace for processing PDF documents. Both Redactwizard Legal 2 and Redactwizard Pro 2 support raster output; native vector redaction is available only in Redactwizard Pro 2.

After opening a PDF, the active page appears in the central canvas. Use the top toolbar for file actions, scanning, saving, page navigation, and zoom controls. You can mark sensitive areas manually or use local AI detection to identify potential entities.

Changes are applied to the output when you save or export the document. Keep the original file as a separate source, and do not assume that edits displayed on the canvas have already been permanently written to the output file.

The integrated OCR pipeline can extract text from many scanned documents. Accuracy depends on scan quality, rotation, contrast, and document complexity. Handwriting is not reliably recognized.

Recommended workflow

Inspect the document first, then choose manual redaction, Assist, Auto, or a combination. Review all detected and manually marked areas before saving, and inspect the resulting file independently.

7. PDF Editor – Inspection, Navigation and View

Use the top toolbar to navigate between pages, zoom in or out, and check the PDF for potential issues. Zooming only changes how the document is displayed on screen; it does not modify the document itself or the intended position of a redaction area. Select 100% to return to the standard view.

Use the Inspect PDF button to start an automatic document inspection. It analyzes the PDF, identifies the document type, and displays information such as the detected page count and classification confidence. It also identifies pages for which OCR review is recommended, for example in scanned documents whose pages may not contain reliably usable text.

The inspection results help you assess whether optical character recognition (OCR) may be needed before further processing. The displayed confidence refers to the document classification and does not indicate the accuracy of OCR recognition.

Recommendation

Zoom in when placing small redaction areas. Then switch to the full-page view to check their position and coverage. For scanned PDFs, also review the OCR recommendations before proceeding with the detection of sensitive information and redaction.

8. PDF Editor – Manual Editing Tools

The left toolbar contains the manual editing tools.

• Rectangular pen: Draws a rectangular redaction area. This is generally the quickest option for ordinary text.
• Eraser: Removes existing marks or areas from the working canvas.
• Brush tools: Mark irregular areas, such as handwritten notes or signatures, where a rectangle would cover too much surrounding content.
• Text tool: Places additional text on a page.
• Bates tool: Adds sequential page or document numbering at the selected position.

The effect of a manual mark depends on the selected output method and the operation performed when the document is saved.

Recommendation

Use rectangles for ordinary text and a suitable brush or polygon for irregular graphic elements. Check the final output rather than relying only on the canvas display.

9. PDF Editor – Colors and Blur

The color selector determines the color used for manual marks. The Blur function applies a visual blur to the selected area. When exported, the effect is incorporated into the generated output according to the selected output workflow.

Canvas edits such as blur, brush marks, and polygons are included in raster output when the document is saved.

Important: Blur is not secure redaction

Blur is not the preferred method for securely removing confidential information. Use an opaque redaction or an appropriate permanent redaction workflow for sensitive content. Blur is more suitable when some visual structure should remain visible.

10. PDF Editor – Polygon and Scissors Tool

The scissors/polygon tool can mark irregular areas such as seals, angled stamps, signatures, and other graphic elements that cannot be covered neatly with a rectangle.

Set points with individual clicks. Click the starting point to close the shape. Complex polygon marks are included in raster processing when the document is saved.

Recommendation

Use polygon selections when a rectangle is unsuitable. For ordinary text, rectangular areas are usually faster and easier to verify.

11. PDF Editor – Blacklist and Whitelist

The Hand-Tap tool lets you select text directly in the PDF and add it to the Blacklist or Whitelist.

1. Select the Hand-Tap tool.
2. Drag a rectangle around the relevant text on the PDF canvas.
3. Release the mouse button to open the context menu.
4. Select Blacklist or Whitelist.
5. Manage the saved entries later in the Rules section.

• Blacklist: Adds a term that should be considered for redaction during automatic detection.
• Whitelist: Adds a term that should be excluded from automatic redaction.

Adding a term to a list does not immediately redact the selected area. Rules are considered during automatic detection and must be checked in context. A term added to one list may be removed from the opposite list according to the application's rule-management behavior.

Recommendation

Use the Blacklist for recurring terms that should normally be redacted and the Whitelist for recurring terms that are incorrectly flagged. Review both lists periodically so that old rules do not affect unrelated documents.

12. PDF Editor – AI Assist, Auto, ROI Selector, and Highlighter

The PDF Editor provides several ways to use local AI-assisted text analysis. The appropriate workflow depends on whether you want to redact detected entities immediately, review individual suggestions, limit detection to a selected area, or highlight entities without redacting them.

Auto

Auto runs entity detection on the available text and automatically adds detected matches as redaction areas.

Auto is useful for recurring document types when you have already evaluated the detection settings and understand how the model handles the material.

Assist

Assist initially displays detected matches as suggestions. You can inspect the suggested areas and accept or remove each suggestion before proceeding.

Use the pen tool to redact the suggested areas. Clicking on the red dashed rectangle will redact this area.

Assist is useful when every potential redaction needs individual review, especially when working with unfamiliar documents, sensitive records, legal files, or text containing ambiguous names and identifiers.

Region of Interest (ROI) Selector

The Region of Interest (ROI) Selector allows you to select a rectangular area of the current PDF page.

After selecting the region, Auto considers detected matches within the selected area and on that page. Text extraction supports embedded PDF text as well as OCR for scanned pages.

The ROI workflow is useful when only a particular section of a page needs processing or when you want to limit detection to a specific area.

An ROI selection does not automatically guarantee complete coverage of the selected region. Text that is not correctly extracted or recognized can still be missed.

Highlighter

The Highlighter uses the same detection process as Auto but highlights detected matches in yellow instead of redacting them.

When exporting the PDF, native PDF highlight annotations with transparency are used so that the underlying text remains readable.

Highlighter is intended for reviewing, locating, or documenting potential sensitive entities. It is not a redaction method and should not be used as proof that information has been removed.

Choosing a workflow

Use Assist when each suggestion needs to be reviewed. Use Auto for tested and repeatable document workflows. Use ROI to limit processing to a selected page region. Use Highlighter when you want to inspect detected entities without redacting them.

Auto Function

Assist Function

Region of Interest (ROI) Selection

ROI Selection Followed by Auto

Highlighter Function

13. PDF Editor – AI Filter

The expandable AI Filter determines which entity classes are considered during automatic detection. Examples include:

• PERSON
• LOCATION
• ORGANIZATION

Additional labels may be available when extra AI models are enabled. Disable labels that are not relevant to the current task. The selection is saved and reused for later analyses. After changing model activation, refresh the list using the Reload icon.

Recommendation

Enable the entity classes relevant to the document type. This can reduce unnecessary matches, but it does not guarantee that every relevant item will be detected.

Entity Classification Limitations

The AI models used to detect personally identifiable information (PII) operate probabilistically. They may misclassify or misinterpret entity classes, for example by identifying a phone number as a date or a name as a location.

This applies in particular to the PII Helper Model, which can optionally be enabled in Redactwizard Pro 2 and serves as the primary model in Redactwizard Legal 2. This model is optimized for compactness and processing speed, which may result in lower detection precision than larger, more resource-intensive models.

Important

Automated redaction and anonymization can produce both false positives and false negatives. Manually review the complete document after automated processing and before final export or distribution.

14. PDF Editor – Available AI Models

Redactwizard Legal 2 includes the integrated PII Primary model. Redactwizard Pro 2 adds further local models that can broaden detection, depending on which models are enabled.

• PII Primary: General personally identifiable and confidential information.
• PII Helper: Structured identifiers and additional detection tasks.
• Medical: Medical and clinical information.
• Legal EU: Legal entities and terminology.

The available labels depend on the active models. More models do not automatically mean better results for every document type. Additional models can introduce false positives and require more processing time.

Recommendation

Start with the primary model and enable additional models when they provide a clear benefit for the material being processed.

15. PDF Editor – Vector and Raster Output

Redactwizard supports two main methods for PDF output. The available options depend on the product edition and settings.

Native Vector Redaction

Native vector redaction is available only in Redactwizard Pro 2.

Where the PDF structure and selected operation allow reliable processing, affected content can be changed within the PDF structure while unaffected content remains in its native form.

Vector output can be useful when preserving selectable text and other native PDF elements is important. However, PDF files vary considerably, and complex structures cannot always be processed reliably. A successful save is not proof that all confidential information has been removed.

Raster Output

With raster output, each page is rendered to an image, and the edits are embedded in the image used to generate the output PDF. The page no longer contains its original text as a normal selectable text layer.

Raster output can reduce the risk of recovering text from the rasterized page through text selection or copying. It does not replace a complete security review of the final file, including document properties and any other content that may remain in the PDF.

• Redactwizard Legal 2: Raster output is the standard method and cannot be disabled.
• Redactwizard Pro 2: Raster output can be enabled or disabled in the relevant settings.

Recommendation

Raster output can be useful for highly sensitive documents or PDFs with complex internal structures. Always inspect the exported file independently, regardless of the output method.

16. PDF Editor – JPEG Quality

When raster output uses JPEG images, JPEG Quality controls the balance between image quality and file size. Higher quality generally produces larger files. Lower quality can reduce file size but may introduce compression artifacts, especially in small text, fine lines, or signatures.

Recommendation

Start with a medium-to-high setting and inspect a representative output page. Avoid settings that make text or fine details difficult to read.

17. PDF Editor – CCITT G4

CCITT Group 4 (G4) is a lossless compression method for 1-bit black-and-white images. It is suitable for monochrome text pages, black-and-white scans, forms, and documents with black text on a white background.

It does not preserve color or grayscale information. If enabled, content that is not pure black and white may be converted to monochrome.

Recommendation

Use CCITT G4 for true black-and-white pages when compact output is important. Use JPEG for color or grayscale pages.

18. PDF Editor – Saving and Verifying Output

When you save a PDF, Redactwizard generates the output according to the selected settings and applied edits. Vector processing may be checked, and raster output may be used as a fallback where applicable if the expected cleanup cannot be confirmed reliably.

A successful save only confirms that the file was written. It does not prove that every sensitive element was detected or removed.

Recommendation

Open the generated PDF in an independent viewer. Check the complete document, select or search text where appropriate, and confirm that the page count and order are correct.

19. PDF Editor – Bates Numbering

The Bates tool adds sequential numbering to a document. The configured prefix is taken from the application settings. For example:

• LAW-0001
• LAW-0002
• LAW-0003

Recommendation

Configure the prefix before starting a larger document set and use a consistent numbering scheme throughout the matter.

20. Handwriting Recognition in Documents

Handwriting recognition example with the PII model from Redactwizard Legal 2: Sample document from philschmid/pdf-samples. Licensed under the Apache License 2.0.

Redactwizard can recognize not only printed text but also handwritten content in scanned documents and images. The locally executed, model-based OCR converts recognized text into machine-readable text, making it available for further processing within the Redactwizard workflow.

Recognition quality depends on factors such as handwriting style, image resolution, contrast, orientation, and document structure. Handwriting recognition may be incomplete or inaccurate, and correct or complete recognition cannot be guaranteed.

Recommendation

Always review the recognized text manually, especially when processing confidential or personally identifiable information. Before using OCR results for redaction, verify that all relevant text has been correctly recognized and that all information requiring redaction has been effectively removed.

Office Editor

21. Office Editor – DOCX and OpenDocument Files

Available only in Redactwizard Pro 2.

The Office Editor works with supported structured Office and OpenDocument files, including:

• DOCX
• ODF
• XLSX
• PPTX

The editor extracts document content from structures such as paragraphs and tables so that it can be analyzed by the local AI. The output may not preserve every detail of a complex or unusual Office file.

Deeply nested content, unusual objects, complex formatting, and embedded elements may not be processed or preserved completely. Always inspect the resulting document visually before releasing it.

Recommendation

Use the Office Editor when retaining the Office format is important. For particularly complex files, consider a PDF workflow only if conversion is appropriate and the final result can be checked carefully.

22. Office Editor – Auto and Assist

Available only in Redactwizard Pro 2.

The Office Editor provides Auto and Assist workflows:

• Auto: Applies detected findings directly as redactions.
• Assist: Presents detected findings for review so individual suggestions can be accepted or removed.

Recommendation

Use Assist for legal documents, reports, or complex tables where each result should be checked before saving.

23. Office Editor – Highlighting and Before/After Review

Available only in Redactwizard Pro 2.

The Highlight function displays the original text with recognized entities highlighted. It helps you review what the AI identified before those findings are applied.

Recommendation

Check the highlighted results immediately after analysis, especially where the meaning of a name, abbreviation, or phrase depends on context.

24. Office Editor – Result View

Available only in Redactwizard Pro 2.

After analysis, the editor displays a processed text representation using the configured replacements. This provides another opportunity to inspect proposed redactions before saving.

Recommendation

Review both the individual replacements and the overall text flow. A result view cannot, on its own, confirm that the final file's entire structure has been cleaned correctly.

25. Office Editor – DOCX Preview

Available only in Redactwizard Pro 2.

The rendered DOCX preview helps you inspect page layout and the visible result before saving.

Recommendation

Pay particular attention to tables, headers, footers, page breaks, images, and complex formatting. Also inspect the saved file in a compatible external application when layout fidelity matters.

26. Office Editor – Search and Manual Redaction

Available only in Redactwizard Pro 2.

The search function locates specified terms independently of AI detection. A matching term can then be reviewed and redacted manually.

This is useful for known names, file numbers, internal references, or recurring terms that need to be processed even if the AI does not identify them.

Recommendation

Use search as a complement to AI detection. Check each result in context before applying a redaction.

27. Office Editor – Blacklist and Whitelist

Blacklist and whitelist entries adapt automatic detection to recurring document types. The feature is available in the Text Editor and, in Redactwizard Pro 2, the Office Editor.

To add a term from the Text Editor, place the cursor over the relevant word or part of it and open the context menu with the right mouse button. Choose the appropriate list action.

• Blacklist: Treats a term as one that should be considered for redaction in future analyses.
• Whitelist: Excludes a term from automatic redaction in future analyses.

Adding a term to one list may remove it from the opposite list. The change affects rule-based detection; it does not necessarily redact the current text immediately.

Recommendation

Use the Blacklist for recurring terms that should normally be removed and the Whitelist for terms that are repeatedly flagged incorrectly. Review the effects after changing a rule.

Document Comparison

28. Document Diff – Compare Text and Documents

Document Diff compares two text versions or the extracted text of two documents. It is useful for reviewing revisions, checking a revised contract against an earlier version, or finding changes between related documents.

The tool provides three modes:

• Text ↔ Text: Enter or paste the original text and the comparison text, or load .txt files into the two editors.
• Document ↔ Document: Select two supported documents and compare their extracted text.
• One document ↔ Multiple documents: Select a reference document and one or more comparison documents. The result selector lets you switch between individual comparisons.

The document modes currently support PDF and TXT files. PDF text is extracted through Redactwizard's existing PDF processing pipeline. Depending on the source PDF, OCR may be needed; image-only pages or poor scans may produce incomplete or inaccurate text.

Office files such as DOCX, XLSX, and PPTX are not direct inputs for Document Diff unless first converted to a supported format.

Using Text Comparison

1. Choose Text ↔ Text.
2. Enter or paste the original version on the left and the revised version on the right, or load a text file into either side.
3. Select Compare in the toolbar.
4. Review the side-by-side result. Added and removed lines are highlighted, and line numbers help locate changes.
5. Use Reset to clear the current inputs, file selections, and results before starting another comparison.

The two input editors have their own scrollbars so that long texts can be reviewed without forcing the entire page to scroll. The result view also provides scrolling for longer comparisons.

Interpreting the Results

• Added: Text present in the comparison version but not in the reference version.
• Removed: Text present in the reference version but not in the comparison version.
• Changed: Lines that differ between the two versions are shown side by side.
• Unchanged: Lines that match are shown without an added/removed highlight.

The comparison is based on extracted text and line-level differences, not a full semantic or legal analysis. A reflowed PDF, different OCR output, or changed line breaks may appear as differences even when the visible wording is similar. Conversely, a comparison result does not prove that every visual or structural difference has been identified.

Recommendation

Check every reported difference in its original document. For scanned PDFs, verify the relevant page visually because OCR errors can affect the comparison. Document Diff does not edit or redact the source files; save or export any separately edited document using its own editor.

Text Editor

29. Text Editor – Basics

The Text Editor is designed for plain .txt files. Depending on the current workflow, it provides:

• Original text and redacted-text views
• Local AI detection
• Assist functions
• Highlighting
• Search
• Undo and redo
• Copying the displayed result

The workflow is similar to text-based processing in the Office Editor, without the need to preserve a complex Office document structure.

Recommendation

Use the Text Editor for plain text files and the PDF Editor for PDF documents. Use Document Diff when the goal is to compare two versions rather than redact a single version.

30. Text Editor – Highlighting

The Highlight function makes recognized entities visible in the original text. It shows locations the AI has identified or proposed for redaction.

Recommendation

Review highlighted names, addresses, identifiers, and similar terms carefully. Context determines whether a match is sensitive and whether additional information needs to be redacted.

AI Models, Filters, and OCR

31. AI Filters in the Editors

AI filters determine which entity classes are considered during automatic detection. They are available in the PDF Editor, Text Editor, and Batch processing. The Office Editor filter is available in Redactwizard Pro 2.

The selection is stored centrally. Additional models can contribute additional labels. After changing model activation, refresh the filter list using the Reload icon.

Recommendation

Use a focused set of labels for recurring document types. Disabling irrelevant classes can simplify review, but do not disable a class merely to hide unwanted results without checking whether it is needed.

32. AI Detection – Limitations

Named Entity Recognition (NER) is statistical. Even when the model runs locally, it may:

• Miss a sensitive entity.
• Mark non-sensitive text as an entity.
• Inherit errors from OCR or text extraction.
• Misinterpret unusual spellings, abbreviations, tables, or formatting.

Recommendation

Treat AI as a detection aid, not as a replacement for a full human review. Manually check document areas where missed information would have serious consequences.

33. Scanned Documents and OCR

PDFs without a native text layer can be processed through the integrated OCR pipeline. The scanned page is interpreted as an image, and recognized text is then available for subsequent analysis.

OCR quality can be affected by poor scan quality, rotated pages, stains, creases, low resolution, weak contrast, or unusual fonts. Handwriting may not be recognized reliably.

Recommendation

For scanned documents, review the rendered page as well as the AI results. Do not rely on OCR alone for small names, reference numbers, handwritten entries, or other critical details.

Batch Processing

34. Batch Jobs

Batch processing is designed for larger collections of PDF documents. Depending on the available batch options, you can select individual files or process directory structures. Jobs run in the background, allowing you to continue using other workspaces while processing continues.

Recommendation

Use Batch for recurring document types after testing the detection settings and output requirements on a small sample.

35. PDF Batch Processing

Select multiple PDF files or a folder for processing. The AI Filter determines which entity classes are considered during automatic detection. Bates numbering can also be applied when configured.

Batch jobs run in the background, leaving other workspaces and editors available.

Recommendation

Batch is well suited to larger collections and familiar document types. For a single document that requires close attention, the PDF Editor can make it easier to control and review each step.

36. Batch – Labels and Rules

Choose the NER labels required for the batch job. Global blacklist and whitelist entries are also taken into account, allowing recurring terms to be handled consistently across multiple documents.

Recommendation

Test the selected combination of labels and rules on representative sample documents before processing a large directory.

37. Batch – Progress and Log

During processing, the Batch window displays progress and information about the files being handled. The log provides details about completed work and any exceptions or errors. Files completed before a job was cancelled remain available.

The result will be saved in the output folder.

Recommendation

After a large batch job, check the log and inspect representative output files. Do not rely only on the number of generated files.

AI Chat

38. AI Chat and External Language Models

Available only in Redactwizard Pro 2.

The AI Chat in Redactwizard Pro 2 connects your working environment to external large language models (LLMs), such as models provided by OpenAI, Anthropic, and Google.

Unlike a conventional cloud workflow that sends documents as entered, Redactwizard Pro 2 provides local pseudonymization before supported AI Chat requests are transmitted.

1. Local preprocessing: The original input and attached documents are processed on your computer before transmission.
2. Local NER masking: An integrated offline named entity recognition (NER) model identifies potential personally identifiable information (PII) and replaces recognized entities with neutral placeholders, such as [PER_1], [ORG_1], or [LOC_1].
3. Filtered transmission: The pseudonymized text is transmitted to the selected AI provider over an encrypted connection.
4. Local de-pseudonymization: When the cloud model responds using placeholders, Redactwizard can restore corresponding original terms locally for display.

Important privacy notice

AI Chat requests are transmitted to external provider APIs. Local pseudonymization reduces exposure of direct identifiers, but automated recognition may miss sensitive information. Context can also reveal identity even when names are replaced. Confirm that external processing is permitted for the relevant client, case, or document before submitting a request.

Step 1: Select a Provider and Enter Your API Key

To use AI Chat, you need an API key from the respective provider. According to the application's configuration, the key is stored locally on your device in config.json.

1. Open AI Chat in Redactwizard Pro 2.
2. Click the gear icon in the bottom toolbar to open the Chat Settings & API window.
3. Select your preferred provider from the dropdown menu:
   • OpenAI
   • Anthropic
   • Google Gemini
4. Enter your API key in the password field.
5. Click Save.
6. Check the status indicator to confirm whether the API key is configured.

You can configure separate API keys for supported providers. When switching providers, the application may retain the previously saved key for each provider.

Step 2: Load Models and Configure the Blacklist

Once an API key has been saved, Redactwizard can retrieve the available models through the provider's API.

1. Click Load Models, or allow the application to load the list after saving.
2. Review the available models.
3. Select the model you want to use. It may also become active in the quick-selection menu in the main window.
4. Maintain your custom blacklist:
   • The Blacklist is located in the lower section of the settings window.
   • Enter internal project names, case numbers, specialized technical terms, or names that should not be sent to the cloud, then click Add Term.
   • During pseudonymization, blacklist terms are intended to receive priority and be replaced with placeholders such as [BLACKLIST_1].
5. Click Close to return to the chat interface.

39. AI Chat – Attach Documents and Review the Preview

Available only in Redactwizard Pro 2.

You can provide supported documents as context for AI Chat. This allows you to summarize contracts, compare clauses, analyze inconsistencies, or prepare documents.

Step 3: Attach a Document

Click the plus icon to the left of the input field to select a file.

Supported file formats include:

• PDF documents (.pdf) with a readable text layer
• Microsoft Word (.docx)
• OpenDocument formats (.odt, .ods, .odp)
• Plain text files (.txt)

Vector PDFs and Scanned Documents

Text extraction in AI Chat is text-based. Digital vector PDFs with selectable text are supported by the text-extraction workflow. Scanned PDFs without a text layer do not contain directly extractable text and may be reported as containing zero characters.

If the source PDF contains only scanned images, run OCR in the main application first, where supported, and verify the extracted text before using it as chat context.

Protecting File Names

For privacy reasons, Redactwizard is designed not to transmit original file names to the cloud API. File names can contain sensitive client or case information. Instead, extracted content is provided under a neutral heading such as --- Extracted Document Text ---.

Step 4: Pseudonymize and Preview Before Sending

Before data is sent to an external provider, you can apply pseudonymization and inspect the content that is intended to be transmitted.

1. Enter your question or instructions in the text field.
2. Click Pseudonymize.
3. The local NER model analyzes your input and attached document.
4. Recognized names, companies, locations, and other supported entities are replaced with placeholders such as [PER_1], [ORG_1], and [LOC_1].
5. The document badge changes to indicate that it has been pseudonymized.
6. Click Preview on the document badge.

A separate preview window displays the extracted and pseudonymized text intended for the cloud request.

Review before transmitting

Check the complete preview carefully. Confirm that names, case numbers, contact details, confidential passages, and other identifying information have been handled appropriately. Pseudonymization is not a guarantee that every sensitive detail has been removed.

Step 5: Inspect the Transmitted Context

Each submitted user message includes an eye icon for inspecting the transmitted context.

Click the icon next to a submitted message to review the request:

The window may display:

1. System Instruction: Instructions supplied by Redactwizard, including directions for handling placeholders.
2. Previous Conversation: Earlier messages in the current session, in the form prepared for transmission.
3. Current User Message: Your instructions together with the pseudonymized document text.

Inspect this information when you need to verify the contents of a request sent to an external provider.

The cloud AI typically responds with pseudonymized placeholders, which Redactwizard replaces with the corresponding entities.

40. AI Chat – De-pseudonymization and Security Limitations

Available only in Redactwizard Pro 2.

Automatic De-pseudonymization of Responses

After the cloud model responds, the response may contain placeholders. Where the corresponding mapping is available, Redactwizard replaces these placeholders locally for display.

For example, the cloud model might receive:

"Under Section 4, [PER_1] is liable to [ORG_1] for claims within the jurisdiction of [LOC_1]..."

The local application may display the corresponding names in the result:

"Under Section 4, Dr. Robert Jenkins is liable to Acme Corp for claims within the jurisdiction of Berlin..."

This process is intended to keep direct identifiers out of the transmitted text while restoring readable output locally. Its effectiveness depends on successful detection, consistent placeholder handling, and the integrity of the local mapping.

The mapping table is intended to remain on the local computer. If the application provides a function to clear the chat, use it when a session is complete, while remembering that this does not delete data already transmitted to a cloud provider or copies retained elsewhere.

1. Limitations of the Local NER Model

• Redactwizard uses a locally executed, quantized Transformer model designed for CPU inference.
• The model can operate without a cloud service or dedicated graphics card.
• Very uncommon surnames, typographical errors, lowercase names, unusual abbreviations, and context-dependent identifiers may escape detection.
• No compact local model can guarantee a 100% detection rate.

Recommendation

Always click Pseudonymize before sending sensitive material and review the complete preview. If an important identifier has not been replaced, edit or remove it manually before transmitting the request.

2. Contextual Re-identification and Quasi-identifiers

Pseudonymization protects direct identifiers, such as names, only when they are correctly detected and replaced. A document may still reveal someone's identity through unique facts, job titles, dates, locations, relationships, or other contextual details.

For example:

"[PER_1] is the chairman of the smartphone manufacturer founded in Cupertino in 1976..."

A language model may be able to infer the person's identity even though the name itself has been replaced.

Recommendation

For highly sensitive matters, remove or generalize unique factual details before sending the text. Review the combination of facts in addition to direct identifiers.

3. Vector PDF Text Quality

The quality of extracted text depends on the source PDF. Poorly generated PDFs may contain merged words, missing spaces, incorrect character encodings, or incomplete text.

Recommendation

Review the pseudonymized preview carefully. If the extracted text is incomplete or inaccurate, correct it manually or use an appropriate OCR workflow before submitting the request.

4. The Blacklist as an Additional Safeguard

Use the blacklist to identify project-specific code names, case numbers, internal references, or unusual terms that the general-purpose NER model may not recognize as sensitive.

Recommendation

Maintain the blacklist for known sensitive terms, but do not treat it as a substitute for inspecting the final pseudonymized text.

Practical Step-by-Step Checklist

1. Configure the provider: Open Settings and enter your API key.
2. Load a document: Attach a supported PDF or DOCX file. For scanned PDFs, perform OCR in the main application first if required.
3. Write your question: Enter your instructions in the input field.
4. Pseudonymize: Click Pseudonymize before sending.
5. Review the preview: Inspect the complete pseudonymized text.
6. Send: Click Send or press Ctrl + Enter.
7. Verify the request: When necessary, click the eye icon to inspect the context prepared for transmission.
8. Clear the session: Use the application's clear or delete function when the conversation and local mapping are no longer needed. Remember that this does not revoke or erase information already transmitted to an external provider.

Clio Integration

41. Clio Integration

The Clio integration can retrieve matters and documents from a connected Clio environment and open them in Redactwizard.

After authorization, select the relevant environment, search for a Matter, and choose the document you need.

Recommendation

Use the integration for recurring Clio workflows to reduce manual file browsing. Verify the selected Matter and document before processing.

42. Clio – Regions and Authentication

Clio authentication uses OAuth 2.0 through the external web browser. Redactwizard does not receive your Clio password as part of this sign-in flow. After authorization, the application can access the data made available by the connected Clio environment.

Select the region that corresponds to your Clio environment, such as Europe, the United States, or Canada, where applicable.

Recommendation

Confirm the region before signing in and verify that the expected Matter and documents appear after authorization.

Archive and Recycle Bin

43. PDF Archive

The internal archive provides optional local storage for PDF documents. Redactwizard uses a local database for archive management. A pdf_archive directory is created alongside the portable application and contains the archive database and associated key material.

Archiving is optional and creates an additional local copy of the document.

Recommendation

Archive only documents for which an additional copy is needed. Include the archive data in your backup and access-control considerations.

44. Viewing and Exporting Archived Documents

Archived PDFs can be managed in the archive. Information such as file name, creation date, and file size may be displayed. Documents can be exported to a location of your choice.

Recommendation

Check the export folder and open the exported file before transferring or releasing it.

45. PDF Archive Recycle Bin

When a document is deleted from the archive, it is first moved to the internal Recycle Bin. Archived objects can be restored from there, or the recycle bin can be emptied to permanently remove the stored archive objects from the application archive.

Recommendation

Empty the recycle bin only after confirming that the archived objects are no longer required. This operation is not a substitute for managing other copies or backups of the same files.

Time Tracking

46. Billing Timer

The Billing Timer records time spent on document work. It can run while you review documents, apply redactions, or perform other processing steps, independently of the active workspace.

Recommendation

Start the timer when the billable task begins and stop it when the relevant work is complete. Check the recorded time before using it for billing.

Rules, Access, and Settings

47. Managing Rules

The Rules section is used to maintain shared terms and application rules, including blacklist and whitelist entries. These rules can help keep processing consistent for recurring document types.

Recommendation

Review rules periodically and remove entries that are no longer relevant. A rule that is useful for one client or document class may cause false positives or missed redactions in another context.

48. PIN Protection

Redactwizard can use a four-digit PIN as an additional layer of local application access control. The PIN is not a replacement for Windows user accounts, file permissions, or device security.

Recommendation

Consider PIN protection on shared or unattended workstations, and secure the Windows account and application data as well.

49. License and Demo Mode

Redactwizard can be started in Demo Mode for evaluation. A valid license key unlocks the features associated with the licensed product edition. Available functionality may differ between Redactwizard Legal 2 and Redactwizard Pro 2.

Recommendation

Before productive use, check the license status and confirm that the active edition includes the features required for your workflow.

50. Application Settings

The settings section controls application behavior. Depending on the edition, available settings include:

• Interface language and color theme
• License and access settings
• AI model activation and entity-processing parameters
• OCR image-processing and performance settings
• PDF rasterization and vector/raster behavior
• JPEG quality and CCITT G4 output
• Bates numbering prefix
• Other application options

Save changes using the designated Save function. Changes may not take effect until they have been saved.

AI Model Settings

Configure which local Named Entity Recognition (NER) models are active for entity detection. The available models and labels depend on the product edition. After changing model activation, refresh the AI filter list using the Reload icon.

Model licenses and attribution are supplied with the applicable model components. Refer to the model documentation distributed with your build for the applicable license details.

Maximum Tokens per Chunk

This setting controls the maximum number of tokens processed in a single NER chunk. If entities are truncated at chunk boundaries or seem to be missed near those boundaries, a smaller chunk size may help.

Smaller chunks can increase processing time and may affect processing efficiency. Test any changes on representative documents before applying them to large collections.

Minimum Score

The minimum score sets the confidence threshold for considering a detected entity. Lowering the threshold may surface more possible entities but can also increase false positives. Raising it may reduce false positives while allowing some sensitive information to go undetected.

Recommendation

Change this setting carefully and review sample results after each adjustment. Do not choose a threshold based only on processing speed or the number of highlights.

NER CPU Threads

This setting controls the number of CPU threads allocated to NER processing. Performance depends on the processor, model, workload, and other applications running on the computer. More threads do not always produce a proportional speed increase.

OCR Settings

Redactwizard uses the local OCR Tensor engine with text-detection and text-recognition models. OCR can detect text regions in page images and attempt to recognize their textual content. The OCR models are loaded locally; this OCR workflow does not require a cloud-based OCR service. The model is only conditionally suitable for detecting signatures.

hese settings control the resolution of the image processed by OCR and the number of CPU threads used by the OCR inference engine.

Maximum OCR Image Edge

This setting controls the maximum length, in pixels, of the longest edge of an image passed to OCR. If the original image exceeds the configured limit, it is resized proportionally before text detection and recognition. Images that are already within the limit are not enlarged.

A smaller value can reduce processing time and memory use, particularly for large scanned pages. However, downscaling can remove fine details and make small text, low-resolution text, or faint characters more difficult to detect or recognize.

A larger value preserves more image detail and may improve recognition of small text, but it can increase processing time and memory use.

The default value is 1,000 pixels. Use representative documents to find a suitable balance between speed and recognition quality.

Recommendation

When processing documents containing handwritten notes, annotations, or signatures, we recommend increasing the maximum edge length (e.g., to 1,500 or 2,000 pixels). Handwriting relies on fine lines, variable pressure, and delicate character strokes that can easily be lost or blurred when images are downscaled. A higher edge length preserves these critical visual details, significantly improving OCR recognition and downstream PII detection accuracy for handwritten text, at the cost of moderately higher processing time and memory usage.

OCR CPU Threads

This setting controls the number of CPU threads made available to the OCR inference engine. The default is based on the number of logical processors reported by the system.

Increasing the thread count may improve processing speed on suitable hardware, but the benefit depends on the processor, image size, model workload, and other applications running at the same time. More threads do not guarantee faster OCR and may increase CPU contention.

The OCR thread setting is separate from the NER CPU Threads setting. Adjust them independently and compare results on representative documents.

OCR Limitations

OCR is intended for detecting and recognizing text. Results depend on image quality, resolution, contrast, font size, orientation, and handwriting.

Although OCR may detect signatures under certain conditions, it is not a dedicated signature-detection system. A signature may be detected if its strokes resemble text, but signatures and other non-standard handwritten marks may be missed or recognized incorrectly. Do not rely on OCR alone to locate every signature or other non-text element that may contain sensitive information.

Always review OCR-generated redaction suggestions before producing the final document, especially when processing scanned pages or documents with small, faint, handwritten, or overlapping text.

PDF Settings

JPEG Quality

Controls the JPEG quality used when generating rasterized PDF content. Higher quality generally preserves more image detail but can increase file size. Lower quality may reduce file size while making small text or fine details less clear.

Always Rasterize Vector PDFs

When enabled, vector PDFs are rendered to page images for output. When this option is disabled, the application may use native vector processing for supported PDFs and operations.

In some circumstances, a visual black box can obscure text without removing the underlying text object. An appearance that looks redacted is therefore not, by itself, proof of secure removal.

PDF files differ in structure, and native vector processing cannot be guaranteed to work reliably for every file. Carefully review the saved document regardless of the selected method.

• Redactwizard Legal 2: Raster output is fixed as the standard output method.
• Redactwizard Pro 2: Raster output can be enabled or disabled in the applicable setting.

CCITT G4 for Black-and-White Raster Output

CCITT G4 can reduce the size of monochrome page images. It uses 1-bit black-and-white output, so color and grayscale information is not preserved and may be converted to monochrome.

Bates Stamp

Bates numbering can be used to identify and organize pages in legal and other document workflows.

Prefix

The prefix is added to the numbering sequence, for example, LAW-0001. Choose a prefix that clearly identifies the relevant case, project, or document set, and keep the numbering consistent.

Access Control

Review the available access-control settings before saving. Application-level controls do not replace Windows account security, filesystem permissions, or an appropriate backup and retention policy.

51. Language and Themes

The interface can be switched dynamically between German and English. Two color themes are available:

• Light
• Dark

The selected theme is applied to the interface, and AI-label colors follow the active theme.

Recommendation

Choose a theme with sufficient contrast for long work sessions. Changing the interface theme does not change AI processing or the content of a PDF export.

Completion and Final Review

52. Saving and Preserving the Original

Saving creates an output file. Keep the original document as an unchanged source, separate from the processed file. For PDFs, redactions and other edits are applied during the save/export process according to the chosen output method.

Recommendation

Use a clear naming scheme for output files and check the destination before saving so that the original is not accidentally overwritten or released.

53. Secure Redaction and Final Verification

A black shape drawn over text does not necessarily remove the underlying content. Depending on how a PDF was created and saved, the text may remain selectable, searchable, or otherwise recoverable. The output method and the structure of the source document both matter.

Redactwizard uses raster output or supported vector processing according to the active edition and settings. Raster output embeds visible page edits into rendered page images, but the entire output file still needs to be reviewed. Do not infer that metadata, attachments, or other non-page content have been removed unless those elements have been checked separately.

Final Review Checklist

Before releasing a sensitive document, verify that:

• All sensitive text passages have been removed or properly obscured.
• Headers, footers, page numbers, captions, and footnotes have been checked.
• Comments, tracked changes, attachments, metadata, and other additional content have been considered where applicable.
• The page count and page order are correct.
• OCR-based documents have been checked visually, including small print and handwritten content.
• Office document layouts, tables, headers, footers, and page breaks are correct.
• The generated file opens correctly in an independent viewer or compatible application.
• No original or unprocessed version is accidentally included in the release package.

For particularly sensitive documents, review every page rather than relying on selected samples or on a successful save operation.

Recommendation

Open the final output outside Redactwizard and perform an independent review before sharing or publishing it.

Recommended Workflow

For a typical PDF redaction task, use the following sequence:

1. Open the document. Check the file type, page count, and visible structure.
2. Choose the detection method. Use manual processing, Assist, Auto, or a combination.
3. Review the AI filter. Enable the entity classes relevant to the task.
4. Review detected findings. Check AI results and add missing areas manually.
5. Apply recurring rules. Use blacklist and whitelist entries where appropriate.
6. Choose the output method. Check vector/raster settings where available. CCITT G4 may reduce file size for black-and-white pages.
7. Save to a separate output file. Preserve the unchanged original.
8. Verify the output independently. Open it in an external PDF viewer and check the complete file.
9. Release only after review. Confirm that the correct output file is being shared.

For comparing versions rather than redacting a document, open Document Diff, select Text ↔ Text, Document ↔ Document, or one document against multiple documents, and review every reported difference in the source files.

Recommendation

For important documents, combine automated detection with manual review, suitable output settings, and independent final verification.

Quick Overview

Important Usage Notice

Redactwizard assists with identifying and removing sensitive information. Automated detection can be incomplete even when all processing is performed locally. This is particularly relevant for unusual documents, OCR output, tables, graphics, complex Office files, and atypical spelling.

Document Diff compares extracted text; it is not a semantic review and does not guarantee that all visual or structural changes have been detected. AI detection and document comparison should both be treated as aids to manual review.

Redactwizard is not a replacement for human review of the final document. For important or sensitive documents, a responsible person should review the complete output before it is released.