LANGCHAIN — Handling PII Data in LangChain

The computer was born to solve problems that did not exist before. — Bill Gates

Handling PII data in LangChain

As the use of AI language models (LLMs) becomes more prevalent, it’s crucial to handle personally identifiable information (PII) data effectively. In this article, we’ll explore how to manage PII data in LangChain, including methods for anonymizing data and ensuring data privacy when using LangSmith for logging conversations.

Why you should care about masking inputs for LLM providers

LLM providers like OpenAI, Anthropic, and Cohere have their own privacy policies detailing how they utilize the data sent to them through their APIs. To ensure data privacy, it’s essential to understand these policies and consider anonymizing PII data before sending it to LLM providers.

Sanitizing data for LLM providers

Microsoft Presidio is a valuable tool for PII identification and anonymization in input text. It uses a combination of methods to identify PII data and then anonymizes it by replacing it with generic equivalents. Here’s an example of how to use Microsoft Presidio with LangChain:

anonymizer = PresidioAnonymizer()
template = """Rewrite this text into an official, short email:
{anonymized_text}"""
prompt = PromptTemplate.from_template(template)
llm = ChatOpenAI(temperature=0)
chain = {"anonymized_text": anonymizer.anonymize} | prompt | llm
response = chain.invoke(text)

OpaquePrompts is another option to anonymize data, using a single ML model to detect and mask PII data. Here’s how to use OpaquePrompts with LangChain:

chain = LLMChain(
  prompt=prompt,
  llm=OpaquePrompts(base_llm=OpenAI()),
  memory=memory,
)

Sanitizing data for LangSmith

When using LangSmith to log app conversations, it’s important to ensure that the data saved is sanitized. One approach is to hide both inputs and outputs from LangSmith using environmental variables:

LANGCHAIN_HIDE_INPUTS=true
LANGCHAIN_HIDE_OUTPUTS=true

Another option is to mask the input before saving it, ensuring that both the LLM and tracing software receive masked inputs.

Conclusion

Effectively handling PII data is crucial for building safe and reliable data applications. By utilizing tools like Microsoft Presidio and OpaquePrompts, and considering options for LangSmith logging, developers can ensure data privacy within the LangChain ecosystem.