BlindLLM

BlindLLM is a Python library that provides a secure way to interact with LLM APIs by automatically anonymizing sensitive information in prompts and de-anonymizing the responses. It supports multiple LLM services including OpenAI, Mistral, and Claude.

Features

• Automatic anonymization of sensitive data (names, emails, organizations, phone numbers, etc.)
• Support for multiple LLM services:
  • OpenAI (GPT-4o, GPT-4, GPT-3.5-turbo)
  • Mistral AI
  • Anthropic's Claude
• Simple, unified interface for all supported services
• Automatic de-anonymization of responses

Installation

pip install .

Requirements

• Python 3.8 or higher
• API keys for the services you plan to use
• text_anonymizer library

Quick Start

from blindllm import BlindLLM

# Initialize with your chosen service
llm = BlindLLM(
    service="openai",  # or "mistral" or "claude"
    model="gpt-4",    # model name for the chosen service
    api_key="your-api-key"
)

# Make a call - get all information including anonymized data
prompt = "John Smith from Acme Corp (email: john.smith@acme.com) asked about the project status."
result = llm.call(prompt)

# Access all available information
print(result['anonymized_prompt'])   # The anonymized version of your prompt
print(result['anonymized_response']) # The raw response from the LLM
print(result['response'])           # The final de-anonymized response
print(result['anonymization_map'])  # The mapping of tokens to original values

Supported Services and Models

OpenAI

• gpt-4o
• gpt-4
• gpt-3.5-turbo
• And other available OpenAI chat models

Mistral

• mistral-large-latest
• mistral-medium-latest
• mistral-small-latest
• mistral-tiny-latest

Claude

• claude-3-5-sonnet-20241022
• claude-3-opus-20240229
• claude-3-sonnet-20240229
• claude-3-haiku-20240307
• And other available Claude models

How It Works

1. Anonymization: When you make a call, BlindLLM uses the text_anonymizer library to anonymize sensitive information in your prompt by replacing it with tokens (e.g., [ENTITY_PERSON_1], [ENTITY_EMAIL_1], [ENTITY_ORG_1]).

2. API Call: The anonymized prompt is sent to the chosen LLM service.

3. De-anonymization: The response from the LLM is automatically de-anonymized by replacing the tokens with the original sensitive information.

Example

from blindllm import BlindLLM

llm = BlindLLM(
    service="openai",
    model="gpt-4",
    api_key="your-openai-api-key"
)

# Example of what happens behind the scenes:

# 1. Your prompt:
prompt = "John Smith from Acme Corp (email: john.smith@acme.com) asked about the project status."

# 2. Anonymized before sending to API:
# result['anonymized_prompt'] will be something like:
# "[ENTITY_PERSON_1] from [ENTITY_ORG_1] (email: [ENTITY_EMAIL_1]) asked about the project status."

# 3. Raw LLM response with anonymized tokens:
# result['anonymized_response'] contains the direct response from the LLM

# 4. Final de-anonymized response:
# result['response'] contains the response with original names restored

# 5. Mapping information:
# result['anonymization_map'] contains:
# {
#   "[ENTITY_PERSON_1]": "John Smith",
#   "[ENTITY_ORG_1]": "Acme Corp",
#   "[ENTITY_EMAIL_1]": "john.smith@acme.com"
# }

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

License

This project is licensed under the MIT License - see the LICENSE file for details.