precept-ai-sdk 0.1.0

Pre-embedding data governance layer for AI systems

Precept SDK

The pre-embedding data governance layer for AI systems.

Precept sits between your raw data and AI memory — enforcing security, privacy, and compliance rules before anything gets embedded into a vector database.

Installation

pip install -e .
python -m spacy download en_core_web_sm

Quick Start

from precept import PreceptGuard

# Initialize with a policy
guard = PreceptGuard(policy='gdpr')

# Sanitize text before embedding
result = guard.sanitize('John Doe email is [e-mail] phone [phone]')

print(result.clean_text)
# Output: 'PERSON_1 email is EMAIL_1 phone PHONE_1'

print(result.detections)
# List of Detection objects showing what was found and replaced

print(result.compliance_status)
# {'gdpr': True}

print(result.tokens_consumed)
# Word count of the processed text

print(result.audit_id)
# UUID for this sanitization event

Policies

Precept ships with pre-built compliance policy presets:

Policy	What It Redacts	Bias Mode	Quality
pii	Names, emails, phones, SSNs, credit cards, IPs	Off	Off
gdpr	All PII + dates + proprietary terms	Flag	On
hipaa	All 18 PHI categories (names, dates, SSNs, emails, phones, etc.)	Remove	On
ndpr	GDPR-equivalent + Nigerian ID patterns (NIN, BVN)	Flag	On

Using a single policy

guard = PreceptGuard(policy='hipaa')
result = guard.sanitize(clinical_note)

Using multiple policies

guard = PreceptGuard(policies=['gdpr', 'hipaa'])
result = guard.sanitize(document)
# result.compliance_status -> {'gdpr': True, 'hipaa': True}

Custom proprietary terms

guard = PreceptGuard(
    policy='gdpr',
    custom_terms=['Project Apollo', 'Operation X', 'AcmeCorp']
)
result = guard.sanitize('The Project Apollo launch is scheduled for Q3.')
# 'Project Apollo' replaced with 'PROPRIETARY_1'

Multi-tenant access control

guard = PreceptGuard(policy='gdpr', tenant_id='customer_123')
result = guard.sanitize(document)
# result.clean_text ends with [PRECEPT_TENANT: <hash>]
# result.tenant_tag contains the 16-char SHA256 hash

Bias detection

GDPR and NDPR policies flag biased sentences; HIPAA removes them:

guard = PreceptGuard(policy='hipaa')
result = guard.sanitize('The patient recovered well. A nurse used a slur.')
# Sentence with slur is removed in HIPAA mode (bias_mode='remove')
print(result.bias_flags)  # List of flagged sentences

Data quality checks

Quality checks run automatically with GDPR, HIPAA, and NDPR policies:

guard = PreceptGuard(policy='gdpr')
result = guard.sanitize('')
print(result.is_safe)           # False — empty text
print(result.quality_report)    # QualityReport with details

API Key (Optional in Phase 1)

If you have a Precept API key, pass it to enable usage tracking:

guard = PreceptGuard(api_key='prct_live_xxxx', policy='gdpr')

Without an API key, the SDK works fully offline — all detection and redaction happens locally.

Running Tests

pip install -e ".[dev]"
python -m spacy download en_core_web_sm
pytest tests/ -v

Project Structure

precept-sdk/
  precept/
    __init__.py          # Public API — exports PreceptGuard
    guard.py             # Core PreceptGuard class
    models.py            # Pydantic data models
    client.py            # HTTP client for API calls
    detectors/
      pii.py             # PII detection (spaCy + regex)
      proprietary.py     # Proprietary term detection
      bias.py            # Bias and toxicity detection
      quality.py         # Data quality checks
      access.py          # Multi-tenant access control
      data/
        bias_terms.txt   # Curated bias term wordlist
    policies/
      base.py            # Base policy class
      gdpr.py            # GDPR policy preset
      hipaa.py           # HIPAA policy preset
      ndpr.py            # NDPR policy preset
  tests/
    test_guard.py        # Guard integration tests
    test_pii.py          # PII detector unit tests
    test_all_modules.py  # Bias, quality, access, full pipeline tests
  setup.py
  README.md

Dependencies

• spacy>=3.7.0 + en_core_web_sm model
• pydantic>=2.0
• httpx>=0.24

Phase 1 — What's Built

• PII detection and redaction (names, emails, phones, SSNs, credit cards, IPs, dates)
• Proprietary term detection with custom term lists
• GDPR, HIPAA, and NDPR compliance policy presets
• Semantic placeholder replacements (PERSON_1, EMAIL_1, etc.)
• Local audit ID generation
• Stub API client (fully connected in Phase 2)

Phase 4 — What's New

• Bias detection: Curated wordlist matching for discriminatory terms with flag/remove modes
• Data quality checks: Empty text, too-short, encoding errors, duplicate detection (session-scoped)
• Multi-tenant access control: SHA256 tenant tagging, isolation verification
• Extended SanitizeResult: is_safe, quality_report, bias_flags, tenant_tag fields
• Updated policy presets: GDPR/NDPR flag bias, HIPAA removes bias, all run quality checks
• Full pipeline: Quality → PII → Proprietary → Bias → Access control in one sanitize() call