👉 Try the demo: https://github.com/LM-Kit/lm-kit-net-samples/tree/main/console_net/document-intelligence/smart-redaction/smart_pii_redaction

Smart Redaction for C# .NET Applications (PII Detection + PDF Redaction)

🎯 Purpose of the Sample

The Smart Redaction demo shows how to combine two LM-Kit.NET capabilities into one compliance-ready workflow: PiiExtraction finds personally identifiable information (PII) in a PDF, and PdfRedactor permanently removes the approved values. A human reviews every suggestion before anything is removed, so accountability stays with a person, not the model. Everything runs 100% on-device: no cloud, no data leaves the machine.

👥 Industry Target Audience

Banking and insurance: redact PII from KYC files, claims, and statements for GDPR reviews.
Healthcare: remove patient identifiers from records before sharing (HIPAA).
Legal and government: prepare sensitive documents for disclosure and FOIA responses.
AI teams: neutralize documents before using them to train or fine-tune models.

🚀 Problem Solved

Drawing a black box over a name in a PDF still ships the name underneath: the text is recoverable by copy-paste, extraction, or raw stream inspection. Manual redaction is also slow and misses occurrences. This sample automates detection and applies true redaction (the underlying data is deleted, not covered), while keeping a human in the loop for the final decision. The result carries no recoverable trace of the removed values.

💻 Sample Application Description

The Smart Redaction demo is a console application that:

Lets you select a model (or paste a custom model URI or model ID)
Redacts a PDF, or an image (converted to a searchable PDF first); ships a bundled sample document
Detects PII across every page with a type, confidence score, and bounding boxes
Presents a human-in-the-loop review where you keep or drop each suggestion
Permanently removes the approved values and writes a redacted copy
Verifies the output with a fresh search, proving the values are gone

✨ Key Features

On-device PII detection: person, email, phone, address, date of birth, SSN, credit card, bank account, and more
Confidence scoring: every suggestion carries a confidence value
Human-in-the-loop review: approve or reject per item, or in bulk (secure default is redact)
True redaction: text glyphs, image pixels, vector graphics, and annotations are deleted, not covered
Precise placement: detected occurrences (TextRegion) feed PdfRedactionArea.FromRegion, so each mark hugs the matched glyphs
Post-redaction verification: confirms the removed values are unrecoverable

Example Output

Detected 8 PII item(s) in 4.2s (overall confidence 0.94).

Review detected PII (human-in-the-loop):
  [Enter]/y = redact    n = keep    a = redact all remaining    k = keep all remaining

  [1/8] Person  "Jane A. Doe"  (confidence 0.98, 2 occurrence(s), page 1)   redact? [Y/n/a/k] a
  ...
Approved 8 of 8 item(s) for redaction.

Redaction complete: 96 glyphs and 0 annotation(s) removed across 1 page(s).
Saved: .../account_application_redacted.pdf

Verifying the redacted output (fresh, cache-free search)...
  All approved values are unrecoverable by text extraction.

🏗️ Architecture

        ┌──────────────┐     ┌───────────────────┐     ┌──────────────┐     ┌──────────────┐
  PDF → │   IDENTIFY   │ →   │  REVIEW (human)   │ →   │    REMOVE    │ →   │    VERIFY    │
        │ PiiExtraction│     │  keep / drop each │     │  PdfRedactor │     │  PdfSearch   │
        └──────────────┘     └───────────────────┘     └──────────────┘     └──────────────┘
         entities +           approved subset          areas + terms         0 matches
         occurrences                                    -> true removal

Identify: PiiExtraction.Extract(attachment) returns PiiExtractedEntity items, each with a type, Value, Confidence, and Occurrences (TextRegion positions).
Review: a person keeps or drops each suggestion. This is the accountability gate.
Remove: approved occurrences become PdfRedactionAreas (via FromRegion), and the value is also added to SearchTerms as a safety net. PdfRedactor deletes the content.
Verify: a fresh PdfSearch over the output confirms nothing survived.

⚙️ Getting Started

Prerequisites

.NET 8.0 SDK or later
A detection model (default qwen3.5:4b, ~3.5 GB VRAM or CPU), downloaded automatically on first run. Larger and vision models are offered for higher accuracy and scanned documents.
Optional: a free community license key from lm-kit.com

Download

git clone https://github.com/LM-Kit/lm-kit-net-samples.git
cd lm-kit-net-samples/console_net/document-intelligence/smart-redaction/smart_pii_redaction

Run

dotnet run -c Release

Press Enter to accept the recommended model, Enter again to use the bundled sample (or enter a path to your own PDF or image), then review each detected item. The redacted PDF is written next to the input as <name>_redacted.pdf.

🔧 Troubleshooting

Occurrences show "unresolved": the value was detected but not mapped to a page region (common with scanned PDFs). The SearchTerms safety net still removes it. For scanned files, attach an OCR engine in Configuration.cs (extractor.OcrEngine = new LMKitOcr();).
Too many or too few detections: tune with domain Guidance, add a PiiEntityDefinition for custom categories, or apply a confidence threshold during review.
Model download is slow the first time: models are cached locally after the first run.

🚀 Extend the Demo

Add a confidence policy: auto-approve items above a threshold and only prompt for the rest.
Add custom PII categories (for example internal case numbers or SWIFT codes).
Batch the pipeline over a folder and write an audit log of what was removed per file.
Swap the interactive review for an approval queue in a web or desktop app.

Table of Contents