4-Layer Detection Pipeline Patterns → Entities → Expressions → Deep Analysis, with automatic merging. Each layer catches what the others miss.
promptShield chains four complementary detection layers. Patterns catch structured data like phone numbers, SSNs, and IBANs with near-perfect precision. Entities use transformer models to identify context-dependent sensitive data like names, organizations, and addresses. Expressions let you flag any specific word or phrase — client names, project codes, internal identifiers. Deep Analysis adds an optional LLM pass for edge cases that the other layers miss. Results from all layers are automatically merged and deduplicated. 14 Entity Detection Models Finds names, organizations, locations, medical terms, and dozens of other context-dependent sensitive data types — the kind pattern matching alone can't catch. Choose from 14 specialized models, or let Auto mode pick the best one. Every model is scanned for malicious code before release.
Select the entity model that fits your use case. Piiranha handles multiple languages in a single pass. The English Comprehensive model detects 54 sensitive data types including medical record numbers, financial accounts, and legal case IDs. The medical model is tuned for clinical documents. Language-specific models deliver higher accuracy for French, German, Spanish, Italian, Portuguese, and Dutch text. Every entity model is also statically scanned for malicious code with picklescan before being published — the build pipeline refuses to ship any model that fails the scan, so what reaches your machine has been verified safe to load. Built-in & Custom Patterns Built-in patterns for SSNs, credit cards, IBANs, phones, business numbers, and more, tailored to 14 countries. Create your own with Simple block-builder or Advanced mode, plus a live test playground.
promptShield ships with dozens of pre-built patterns for common sensitive data types, localized for 14 countries. Need something custom? Simple mode provides a visual block-builder where you define patterns without writing any code. Advanced mode gives you full pattern syntax with capture groups. Both modes include a live test playground where you can paste sample text and see matches highlighted in real time before applying to your documents. Run a local GGUF model on your GPU, or connect to any OpenAI-compatible API (OpenAI, Claude, Groq, Mistral, Ollama, vLLM) for deep context-aware analysis.
For maximum detection accuracy, add a Deep Analysis layer to the pipeline. Run a GGUF model locally on your GPU for fully offline processing, or connect to any OpenAI-compatible API endpoint. The model analyzes text in context, catching subtle sensitive data that Patterns and Entities might miss — implied references, nicknames, or domain-specific identifiers. Configure temperature, max tokens, and system prompts to tune behavior. Enter specific words or phrases to find, paste from Excel or CSV. Each match can be set to Encode or Remove, with adjustable fuzziness.
The Expression Grid lets you specify exact terms you want detected, perfect for client names, project codes, or internal designators that automatic detection might miss. Paste directly from Excel or CSV for bulk entry. Each term can be set to Encode (reversible) or Remove (permanent redaction), and the fuzziness slider controls how closely text must match. This is your safety net for domain-specific data that only you know about. Tune detection aggressiveness with a fuzziness slider, per-type toggles, and a region grouping control for how nearby words merge.
Fine-tune detection to your exact needs. The fuzziness slider controls how aggressively promptShield expands matches to surrounding context. Per-type toggles let you enable or disable specific sensitive data categories (e.g. detect phone numbers but skip organization names). The region grouping control determines how nearby detected words merge into single regions, important for multi-word names and addresses. 14 Countries, 7 Languages Country-specific patterns for the US, UK, Canada, Australia, France, Germany, Spain, Mexico, Italy, Netherlands, Belgium, Portugal, Brazil, and Switzerland. including business numbers, national IDs, and tax identifiers unique to each country.
Each country profile includes localized patterns for national IDs (SSN, NI number, SIN, TFN, NIR, etc.), tax identifiers, business registration numbers, postal codes, phone formats, healthcare numbers, and more. When you select a country profile, the detection engine is configured with the right patterns. Auto mode detects the likely country from document content and activates the matching profile. Detected sensitive data automatically propagates across all pages, including partial org-name matching and type unification for consistency.
When promptShield detects a name on page 1, it automatically finds and marks every occurrence of that name on all other pages, even with slight variations (e.g. 'Smith & Associates' vs. 'Smith & Assoc.'). Type unification ensures that if an entity is classified as a PERSON on one page, it maintains that classification everywhere. This guarantees consistent encoding across the entire document. Pointer, lasso, and draw tools for selecting and creating regions. Merge, split, re-label, or delete with full undo/redo support.
Interactive tools let you visually manage detected regions on the document canvas. Use the pointer to select existing regions, the lasso to select multiple regions at once, or the draw tool to manually mark areas the AI missed. Right-click any region to merge, split, re-classify the sensitive data type, or delete. Every action supports undo/redo, so you can experiment freely without losing work. Save and load named detection configurations (fuzziness, enabled types, expressions) to reuse across documents.
Create detection templates for recurring document types. A template captures your entire detection configuration: fuzziness level, enabled sensitive data types, expressions, custom patterns, and entity model selection. Save templates with descriptive names like 'Legal Contracts' or 'Medical Records' and apply them instantly to new documents. This eliminates repetitive setup and ensures consistent processing across your team. Choose per-region: encode (reversible replacement) or remove (permanent black-box redaction). Apply in bulk or one at a time.
Not all sensitive data needs the same treatment. Encoding replaces data with reversible codes so you can restore it later after AI processing. Removal permanently redacts the data with a black box, ideal for information that should never be recoverable. Set the default per sensitive data type, or override individual regions. Bulk selection lets you change the mode for dozens of regions at once. Noise Filtering & Confidence Built-in noise filters reduce false positives for ORG, Location, Person, and Address types. Every detection includes a confidence score.
Entity detection can sometimes be overly aggressive, flagging common words as organizations or generic locations as sensitive data. promptShield includes built-in noise filters that suppress common false positives for each entity type. Every detection also carries a confidence score, letting you quickly identify uncertain matches that deserve manual review. The result is cleaner detection with fewer false positives out of the box. Metadata & Hidden Data Suppression All document metadata, hidden fields, tracked changes, embedded objects, and invisible content are automatically stripped during processing. ensuring no sensitive data leaks through non-visible channels.
Documents often contain hidden sensitive data that isn't visible on the page: author names in metadata, GPS coordinates in images, revision history with tracked changes, comments, hidden text fields, and embedded objects. promptShield automatically strips all of these during processing. This prevents the common scenario where a user carefully redacts visible text but forgets about metadata that reveals the same information. Shortcuts for Power Users Review and protect documents without touching the mouse. Switch tools, act on regions, navigate pages, and undo — all from the keyboard.
Tool switching: V (pointer), S (lasso), D (draw), Space (toggle last tool). Region actions: T (encode selected), D or Backspace (remove selected), Delete (delete region or batch-delete). Navigation: Tab / Shift+Tab to cycle through pending regions, Left / Right arrows to change page. Zoom: + / - / 0 (reset). Selection: Ctrl+A to select all regions on the page. Clipboard: Ctrl+C / Ctrl+V to copy and paste regions across pages. Undo/Redo: Ctrl+Z / Ctrl+Shift+Z. Detection panel: Ctrl+F. All shortcuts are suppressed inside text inputs so they never interfere with typing.