PII Configuration

Supported entity types

NoPII detects 11 entity types by default. Each can be individually enabled or disabled in the admin console.

Identity

Government IDs

India

Off by default. Switch them on under PII Detection in the admin console. Numbers written in Devanagari or other Indian digits are detected too.

Other countries (beta)

Off by default. Switch them on under PII Detection in the admin console. Where a type says "after" a label, it is detected only when that label comes shortly before the number, as in "My TFN is 123 456 782". Ordinary order, invoice and phone numbers pass those types' check digits too often for the number alone to be enough.

Financial

Technical

CREDENTIAL entity type

The CREDENTIAL entity type detects API keys, tokens, secrets, and connection strings in your prompts. This prevents accidental exposure of credentials to LLM providers.

Detected patterns

Named patterns (AWS, OpenAI, Stripe, etc.) are detected with high confidence (0.85–0.95). Generic patterns like Azure SAS tokens start at lower confidence but are boosted when surrounding context words like "key", "secret", "token", or "credential" are present.

Enhanced LOCATION detection

In addition to standard location detection (cities, states, countries), NoPII includes enhanced street address detection that catches US and UK-style addresses with:

• House numbers and street names with common street types (Ave, Blvd, St, Dr, Ln, etc.)
• Directional suffixes (N, S, E, W, NE, NW, SE, SW)
• Unit designators (Apt, Suite, Unit, Floor, Room, #)

Examples: "742 Evergreen Terrace", "456 Oak Avenue, Apt 3B", "123 Main St NW Suite 100"

In Indian addresses, in Hindi or English, house, flat, plot, lane and sector numbers are detected too ("मकान नंबर 12", "Flat 304", "Sector 21"), and so is the six-digit PIN code when an address, a PIN label or a state or city name comes before it. A six-digit number on its own, such as an OTP or an order number, is left alone.

Language support

Pattern-based types (email, card numbers, phone numbers, the identifier types listed above, credentials) are detected in most languages, including messages that mix languages, numbers typed in non-Latin or full-width digits, and values written with no space before or after them in Chinese, Japanese, Korean or Thai text. Names and places are detected in English, and in Hindi written in Devanagari.

Current limits:

• Phone numbers without a country code are recognized for the US, UK, Canada, India, Germany, France, Brazil and Israel. Numbers from other countries are recognized in international format, such as +34 612 34 56 78.
• Names written in Chinese, Japanese, Korean or Thai characters are not yet reliably detected.

Hindi (beta)

With PERSON and LOCATION enabled, names and places written in Devanagari are detected and tokenized like English ones, including in messages that mix Hindi and English. No setting is needed. For example, in "मेरा नाम प्रिया वर्मा है और मैं गोरखपुर से हूँ", both "प्रिया वर्मा" and "गोरखपुर" are replaced with tokens.

Not yet covered: Hindi written in Latin letters ("mera naam Priya hai"), where detection falls back to English rules, and Urdu.

Confidence threshold

The confidence threshold controls how aggressively NoPII detects PII. Each potential PII entity is assigned a confidence score between 0.0 and 1.0:

You can adjust the threshold in the admin console. Use the Detection Preview tool to test different thresholds against sample text before changing your production settings.

Context phrase neutralization

Some LLMs may refuse requests that mention sensitive data categories by name (e.g., "social security number"). NoPII automatically replaces these category phrases with neutral alternatives before sending to the LLM, then restores them in the response.

Built-in replacements

These replacements are case-insensitive and handle singular/plural forms automatically. They apply in addition to PII tokenization - the actual PII values are still replaced with tokens.

Only the "... number" forms are replaced. Bare category names such as "credit card", "driver's license" and "date of birth" are sent as written: their neutral forms would be everyday words, and restoring those would also rewrite places where the model used the word itself.

Custom context replacements

You can add your own phrase replacements in the admin console. This is useful for domain-specific terminology that might trigger LLM safety filters. For example, you could replace "medical record number" with "file reference" or "insurance policy number" with "account reference".

Custom replacements are applied after the built-in ones. Original phrasing is automatically restored in the response. Phrases match whole words only, and can be in any language written with spaces between words, Hindi included. In Chinese, Japanese and Thai, a phrase matches only where spaces or punctuation set it off.

Restoration replaces every occurrence of the neutral term in the response, including any the model wrote on its own. Choose a term the model would not use unprompted: "file reference" is safe, "file" is not. A term that matches a token label (NAME, DATE, EMAIL, PHONE, LOCATION, IDENTIFIER, ADDRESS, GROUP, REFERENCE) is rejected, because it would corrupt the token wrappers in the response.

Managing settings

All PII detection settings are configured through the admin console. You can enable or disable entity types, adjust the confidence threshold, and manage context phrase replacements. Changes take effect immediately - the PII detector cache is invalidated on update. The console lists only the entity types NoPII can detect, and flags any enabled type it cannot.

Detection preview

The admin console includes a detection preview tool that lets you test your PII settings against sample text before changing production configuration. It runs PII detection with your current settings and shows the detected entities with their confidence scores - without tokenizing or forwarding anything.