Analyzer Guide -- Finding Personal Information

Last Updated: 2026-02-11

The Analyzer is the detection engine at the core of cloak.business. It scans text for personally identifiable information (PII) using a combination of pattern matching and natural language processing (NLP) models.

Table of Contents#

1. What the Analyzer Does
2. How Detection Works
3. Confidence Scores
4. Context Words
5. Entity Types
6. Language Selection
7. Tips for Best Results

What the Analyzer Does#

The Analyzer examines text and identifies segments that contain personal or sensitive information. For each detected entity, it returns:

• The entity type (e.g., PERSON, CREDIT_CARD, DE_TAX_ID)
• The exact position in the text (start and end offsets)
• A confidence score indicating how certain the detection is

The Analyzer does not modify text. It only identifies PII. Anonymization is a separate step that uses the Analyzer's output.

How Detection Works#

Detection uses two complementary approaches that run in parallel:

Pattern-Based Recognition (317 Recognizers)#

• Regex patterns match structured data formats: national ID numbers, tax IDs, credit card numbers, phone numbers, IBANs, license plates, postal codes, and more.
• Each recognizer is tuned for a specific country or format. For example, the German ID card recognizer matches the exact format of a German Personalausweis number.
• Pattern recognizers work regardless of the text's language because they match data format, not language.
• Coverage spans 75+ countries and 390+ entity types.

NLP-Based Recognition#

• Named Entity Recognition (NER) models detect names of people, locations, and organizations.
• NLP models understand context and grammar to identify entities that do not follow a fixed pattern.
• Multiple NLP engines are available:
  • spaCy models for 25 languages
  • Stanza NER for 7 languages
  • XLM-RoBERTa transformers for 16 languages

Both approaches run simultaneously. When both a pattern and an NLP model detect the same entity, the higher confidence score is used.

Confidence Scores#

Every detection includes a confidence score from 0.0 to 1.0.

What Affects Confidence#

• Checksum validation: Entity types like credit cards and IBANs include mathematical validation (Luhn algorithm, IBAN check digits). Passing validation significantly boosts confidence.
• Context words: The presence of related words near the detected value increases confidence (see below).
• Pattern specificity: A highly specific pattern (e.g., a German tax ID with exact format) scores higher than a generic numeric pattern.

Context Words#

Each recognizer has language-specific context words that boost detection confidence when they appear near a potential entity.

How Context Words Work#

1. The pattern recognizer finds a value matching a known format.
2. The system checks the surrounding text (typically within a 5-word window) for context words.
3. If context words are present, the confidence score is increased.

Examples#

Context words are defined per recognizer and per language. The system supports context words in all 48 interface languages.

Entity Types#

Detected entities are organized into the following categories:

Personal Identifiers#

Names, dates of birth, ages, genders, nationalities, and biometric identifiers. Detected primarily by NLP models.

Government-Issued IDs#

National ID numbers, passport numbers, driver license numbers, social security numbers, and tax IDs. Covers 75+ countries with country-specific formats.

Financial Identifiers#

Credit card numbers, IBANs, SWIFT/BIC codes, bank account numbers, and securities identifiers (CUSIP, ISIN, SEDOL, LEI).

Location Data#

Addresses, postal codes, GPS coordinates, and IP addresses.

Digital Identifiers#

Email addresses, phone numbers, URLs, MAC addresses, and license plates.

Technical Secrets#

API keys, access tokens, SSH keys, database connection strings, and credentials from 30+ platforms (cloud providers, AI services, SaaS tools).

Healthcare#

Medical record numbers, prescription numbers, diagnosis codes (ICD-10, ICD-11), procedure codes, and country-specific health insurance numbers.

Organization#

Company names, legal entity identifiers, and registration numbers.

Temporal#

Dates, times, and durations that may be personally identifying.

For the complete list of all 390+ entity types and 157+ presets, see the Entity & Preset Inventory.

Backend Request Limits#

The analyzer enforces server-side limits on every request to prevent resource exhaustion. These limits apply regardless of which client sends the request.

Requests exceeding these limits receive a 422 validation error with a descriptive message. Clients can discover the current limits via the GET /limits endpoint.

Language Selection#

Auto-Detect (Default)#

The system identifies the primary language of the input text automatically. This works well for most cases.

Manual Selection#

Choose a specific language to:

• Load the correct NLP model for that language, improving name and location detection.
• Activate language-specific context words for pattern recognizers.
• Improve OCR accuracy when processing images.

What Language Affects#

Tips for Best Results#

1. Provide more context. Longer text gives the Analyzer more surrounding words to evaluate. A credit card number in isolation may score lower than one preceded by "Card Number:".

2. Use the right preset. Country-specific presets enable only the relevant entity types, reducing false positives. If you are processing German documents, select the Germany preset.

3. Check low-confidence detections. Entities with scores below 0.5 are more likely to be false positives. Review them before anonymizing.

4. Select the correct language. If auto-detect picks the wrong language, manually select the correct one. This improves NLP detection significantly.

5. Include headers and labels. Documents with clear labels like "Name:", "Address:", "Tax ID:" provide strong context words that boost confidence scores.

6. Review before anonymizing. Always review detected entities and deselect any false positives before proceeding to anonymization.