FastPII

---

Why FastPII

Most PII tools are built for generic text processing. FastPII is built for AI workflows.

Modern applications increasingly send documents, prompts, support tickets, contracts, medical records, and business data directly into LLMs and AI systems. FastPII acts as the privacy layer between your data and your AI.

Without FastPII          With FastPII

Document                 Document
  ↓                        ↓
LLM                      FastPII
                           ↓
                         LLM

Sensitive data exposed   Sensitive data protected

European-Native Detection

FastPII understands European identifiers and validation rules across 4 countries:

• Czech Republic: Rodné číslo · IČO · DIČ · Bank Accounts · Phone · PSČ · Addresses · Names · Vehicle Plates · Dates of Birth · Email
• Poland: PESEL · NIP · REGON · Postal Codes · Phone · Addresses
• Germany: Steuer-ID · USt-IdNr · Handelsregister · PLZ · Phone · Addresses
• France: SIREN · SIRET · INSEE/NIR · Code Postal · Phone · Addresses

Checksum Validation

FastPII validates identifiers instead of relying solely on pattern matching. Rodné číslo (Mod 11), IČO (weighted Mod 11), PESEL (Mod 10), NIP (weighted Mod 11), REGON (weighted Mod 11), Steuer-ID (ISO 7064 MOD 11,10), USt-IdNr (MOD 11,10), SIREN/SIRET (Luhn), INSEE/NIR (Mod-97), and DIČ formats all require valid checksums — structurally invalid identifiers are rejected before classification.

Context-Aware Detection

Detection is not based on regex alone. Phone numbers require context words or a country-specific prefix (+420, +48, +49, +33). Postal codes require address proximity or labels. Dates of birth need birth-related keywords nearby. Addresses use component scoring. This significantly reduces false positives.

Fully Configurable

Every detector value is a class-level attribute. Confidence thresholds, scoring weights, context words, city lists, and overlap priority are all explicit and overridable — no hidden defaults in the core engine.

Built for AI Workflows

FastPII integrates directly into RAG pipelines, LangChain applications, MCP servers, AI agents, FastAPI applications, and enterprise AI systems.

---

Features

Detection	Identify sensitive Czech and European data
Validation	Validate identifiers using official checksum rules
Privacy Protection	Four modes: anonymize, redact, mask, remove
Explicit Engine	No implicit behavior — you configure, it executes
Framework-Independent SDK	Use as a standalone Python package
Integrations	FastAPI, LangChain, MCP, CLI
Local First	No cloud, no LLM, no external API calls required

---

Installation

pip install fastpii

---

Quick Start

Explicit Engine API (Recommended)

from fastpii import FastPII
from fastpii.core.confidence import ConfidenceScorer
from fastpii.countries.cz import CzechPack
from fastpii.countries.pl import PolishPack

priority = {"rodne_cislo": 100, "pesel": 100, "email": 70, "name": 50, "phone": 20}

engine = FastPII(
    priority=priority,
    confidence_scorer=ConfidenceScorer(
        base_scores={"checksum_validated": 1.0, "context_match": 0.95, "pattern_match": 0.85},
        context_boost=0.10,
    ),
)
engine.register(CzechPack())
engine.register(PolishPack())

text = "Jan Novák, RČ: 800101/1238, IČO: 25596641"
result = engine.detect(text)

for finding in result.findings:
    print(f"{finding.type}: {finding.value}")

Convenience API (PrivacyGuard)

from fastpii import PrivacyGuard

# Single-region detection
guard = PrivacyGuard(regions=["cz"])
result = guard.detect("Jan Novák, RČ: 800101/1238, IČO: 25596641")

# Multi-region detection
guard = PrivacyGuard(regions=["cz", "pl", "de", "fr"])

Note: PrivacyGuard() without regions emits a DeprecationWarning. Explicit region specification is recommended.

Region Quick Start

# Poland
guard = PrivacyGuard(regions=["pl"])
result = guard.detect("PESEL: 44051401458, NIP: 5260250274")

# Germany
guard = PrivacyGuard(regions=["de"])
result = guard.detect("Steuer-ID: 86095742719, USt-IdNr: DE136695976")

# France
guard = PrivacyGuard(regions=["fr"])
result = guard.detect("SIREN: 552120222, SIRET: 73282932000074")

---

Explicit Engine API

FastPII follows an explicit execution model: Core executes, Platform decides. The engine never makes implicit choices.

Required Configuration

• priority — Overlap resolution order (required, no default)
• confidence_scorer — Confidence scoring configuration (optional, no hidden defaults)

from fastpii import FastPII
from fastpii.countries.cz import CzechPack

# Minimal setup — define your own priority dict
priority = {"rodne_cislo": 100, "email": 70, "name": 50, "phone": 20}
engine = FastPII(priority=priority)
engine.register(CzechPack())
result = engine.detect(text)

Overridable Detector Values

Every detector exposes class-level attributes for all configurable values:

from fastpii.detectors.cz.address import CzechAddressDetector

# Override scoring thresholds
class CustomAddressDetector(CzechAddressDetector):
    MIN_ADDRESS_SCORE = 0.3
    SCORE_STREET = 0.4
    SCORE_NUMBER = 0.3
    SCORE_POSTAL = 0.2
    SCORE_CITY = 0.1

# Override context words
class CustomPhoneDetector(CzechPhoneDetector):
    CONTEXT_WORDS = {"tel", "phone", "call"}
    CONTEXT_WINDOW_SIZE = 50

Convenience Defaults

For backward compatibility, PrivacyGuard provides built-in convenience defaults for overlap priority and confidence scoring. These are used internally by PrivacyGuard — for production use with FastPII, define your own priority and scoring configuration based on your domain requirements.

---

Privacy Modes

Anonymize — Replace with [REDACTED]

engine.anonymize("Jan Novák, RČ: 800101/1238")
# → "[REDACTED], RČ: [REDACTED]"

Redact — Replace with PII type label

engine.redact("Jan Novák, RČ: 800101/1238")
# → "[NAME], RČ: [RODNE_CISLO]"

Mask — Replace with asterisks

engine.mask("Jan Novák")
# → "*********"

Remove — Delete PII entirely

engine.remove("Jan Novák")
# → ""

---

AI Use Cases

Protect RAG Pipelines

safe_document = engine.anonymize(document)
embeddings = embed_model.embed(safe_document)

Protect LLM Prompts

safe_prompt = engine.anonymize(prompt)
response = llm.invoke(safe_prompt)

Protect MCP Tools

safe_input = engine.anonymize(user_input)
result = tool.execute(safe_input)

---

Supported Entities

Czech Republic (CZ)

Entity	Detection Method	Checksum
Rodné číslo	Mod 11 checksum + date validation	✓
IČO	Weighted Mod 11 checksum	✓
DIČ	Multi-format + IČO validation	✓
Bank Account	Two-part Mod 11 checksum	✓
Postal Code	Context-gated (PSČ label, city, address proximity)	—
Phone Number	Context-gated (+420 prefix or context words)	—
Date of Birth	Context-gated (birth keywords, intervening date blocking)	—
Address	Component scoring (street + number + city + postal)	—
Name	Czech name dictionary + gender classification	—
Email	Czech TLD detection, markdown mailto handling	—
Vehicle Plate	Regional code validation	—

Poland (PL)

Entity	Detection Method	Checksum
PESEL	Mod 10 checksum + date/gender extraction	✓
NIP	Weighted Mod 11 checksum	✓
REGON	Weighted Mod 11 checksum (9-digit and 14-digit)	✓
Postal Code	Context-gated (DD-DDD format)	—
Phone Number	Context-gated (+48 prefix or context words)	—
Address	Component scoring (street + number + city + postal)	—

Germany (DE)

Entity	Detection Method	Checksum
Steuer-ID	ISO 7064 MOD 11,10 checksum	✓
USt-IdNr	MOD 11,10 checksum (DE + 9 digits)	✓
Handelsregister	Format validation (Court + HRA/HRB/PR + number)	—
Postal Code	Context-gated (5-digit PLZ)	—
Phone Number	Context-gated (+49 prefix or context words)	—
Address	Component scoring (Straße + number + PLZ + city)	—

France (FR)

Entity	Detection Method	Checksum
SIREN	Luhn checksum (9-digit)	✓
SIRET	Luhn checksum (14-digit, embeds SIREN)	✓
INSEE/NIR	Mod-97 checksum (15-char, Corsica 2A/2B handling)	✓
Postal Code	Context-gated (5-digit code postal)	—
Phone Number	Context-gated (+33 prefix or context words)	—
Address	Component scoring (number + Rue/Avenue/Boulevard + city)	—

---

Architecture

FastPII follows an Open Core architecture with a strict separation of concerns:

• Core executes — deterministic detection, validation, redaction
• Platform decides — auto-detection, pack selection, privacy presets
• Enterprise governs — org-level policy, audit, compliance

The OSS core is fully self-hostable, local-only, and explicit. No implicit behavior, no hidden defaults, no auto-detection.

See ADR 0001 for the full boundary definition.

---

Benchmarks

Evaluated on Czech-focused datasets containing contracts, medical records, business registries, support tickets, and adversarial false-positive scenarios.

v0.4.0 overall:

Metric	Score
Precision	84.2%
Recall	80.0%
F1	82.1%

Per-detector:

Detector	Precision	Recall	Notes
IČO	100%	100%	Checksum-validated, no FPs
DIČ	100%	100%	Multi-format detection
Email	100%	100%	Markdown mailto handled
Date	100%	100%	Non-birth dates detected separately
Phone	100%	100%	Context or +420 prefix required
Vehicle Plate	100%	100%	Regional code validation
Date of Birth	100%	86%	Context-gated; rejects generic dates
Postal Code	100%	71%	Subsumed by address in overlaps
Name	80%	100%	Dict-matched; corporate name FPs
Address	71%	63%	Component scoring; partial matches
Rodné číslo	67%	50%	Invalid checksums correctly rejected
Bank Account	100%	0%	Requires labeled context (v0.2.5)

---

Roadmap

Current — Explicit Engine API, Configurable Detectors, CZ/PL/DE/FR Country Packs, Validation Engine, CLI, FastAPI Integration, LangChain Integration

Next — Strict Mode, MCP Integration, RAG Middleware, Improved Address & Bank Account Detection, Additional European Regions

---

Documentation

• Quick Start
• Explicit Engine API Design
• OSS Core Boundary ADR
• OSS / Platform Split Plan
• Migration Roadmap
• Implementation Backlog
• Repo Split Checklist

---

Contributing

Contributions welcome! See Contributing Guide.

Before contributing, review the OSS Core Boundary to ensure your change belongs in the OSS core. The core executes — it does not decide.

---

License

Apache 2.0 — See LICENSE for details.

---