LLMShield v2.1.0 - Enhanced Provider Support and PII Allowlist

This release adds three new LLM providers and an allowlist API for fine-grained PII control.

View this release on PyPI.

New Providers

• Google GenAI - Full support for Gemini models, including function calls and tool results
• Cohere - Support for Cohere chat API with tool call uncloaking
• xAI - Support for xAI/Grok SDK response formats

New Features

• Allowlist API - Exclude specific terms from PII detection via LLMShield(allowlist=["Alexa"]) or per-call with shield.ask(allowlist=["Acme"])
• Python 3.14 Support - Tested and verified across Python 3.12, 3.13, and 3.14

Improvements

• Test Suite Overhaul - Deduplicated and added parameterised tests across all modules.
• CI Modernisation - Upgraded all workflows to uv for development.

LLMShield v2.0.0 - New API and Improved Accuracy

This release makes major improvements to reducing false positives across all entity classes and introduces a more flexible chaining API.

View this release on PyPI.

Improvements

• Reduced False Positives - Overhaul of detection logic to prevent cloaking common words
• Opt-In Concept Detection - CONCEPT detection now disabled by default for technical acronyms
• Chaining API - New without_* methods for specific PII configuration
• Scalable Caching - Default max_cache_size increased to 10,000 message turns
• State-Preserving Chaining - Methods preserve configurations instead of resetting them

LLMShield v1.0.0 - Initial Production Release

We're excited to announce the first stable release of LLMShield, a production-ready, zero-dependency Python library for protecting sensitive information in LLM interactions.

View this release on PyPI.

Key Features

Core Functionality

• Zero Dependencies - Pure Python implementation requiring no external packages
• Universal LLM Compatibility - Works with OpenAI, Anthropic, xAI, and any OpenAI-compatible provider
• Automatic PII Detection - Multi-layered detection system for 9 entity types
• Streaming Support - Real-time processing for streaming LLM responses
• Conversation Memory - Maintains entity consistency across multi-turn conversations
• Selective Protection - Fine-grained control over which entity types to protect

Entity Detection Types

• PERSON - Names and titles (John Doe, Dr. Smith)
• ORGANISATION - Companies and institutions (Microsoft, NHS)
• EMAIL - Email addresses with validation
• PHONE - International phone number formats
• CREDIT_CARD - Credit card numbers with Luhn validation
• PLACE - Locations and addresses
• URL - Web addresses
• IP_ADDRESS - IPv4 addresses
• CONCEPT - Uppercase technical terms (API, SQL)

Provider Support

Full support includes text chat completions, image inputs (for supported models), structured outputs, and tool calling.

• OpenAI - Full support
• Anthropic - Full support
• xAI Grok - Full support
• OpenAI-Compatible - Full support
• Google - Unsupported
• Cohere - Unsupported

Advanced Features

• LRU Cache System - Configurable caching for optimal performance
• Custom Delimiters - Flexible placeholder formatting
• Tool Call Protection - Automatic PII protection in function/tool calls
• High Performance - O(1) entity lookups, 99.9% faster initialization
• Thread-Safe - Singleton pattern with proper locking
• Comprehensive Testing - 98.42% code coverage with 707 tests