When building chat applications, it’s crucial to ensure the secure handling of sensitive data, especially Personal Identifiable Information (PII). PII can be directly or indirectly linked to an individual, making it essential to protect user privacy by preventing the transmission of such data to language models.

​
Example of PII

Consider the text below, where PII has been highlighted:
Hello, my name is John and I live in New York. My credit card number is 3782-8224-6310-005 and my phone number is (212) 688-5500.
And here is the anonymized version:
Hello, my name is <PERSON> and I live in <LOCATION>. My credit card number is <CREDIT_CARD> and my phone number is <PHONE_NUMBER>.

​
Analyze and anonymize data

Integrate Microsoft Presidio for robust data sanitization in your Chainlit application.
Code Example
import chainlit as cl

@cl.on_message
async def main(message: cl.Message):
    # Notice that the message is passed as is
    response = await cl.Message(
        content=f"Received: {message.content}",
    ).send()
Before proceeding, ensure that the Python packages required for PII analysis and anonymization are installed. Run the following commands in your terminal to install them: 
pip install presidio-analyzer presidio-anonymizer spacy
python -m spacy download en_core_web_lg
Create an async context manager that utilizes the Presidio Analyzer to inspect the incoming text for any PII. This context manager can be included in your main function to scrutinize messages before they are processed. When PII is detected, you should present the user with the option to either continue or cancel the operation. Use Chainlit’s messaging system to accomplish this.
Code Example
from presidio_analyzer import AnalyzerEngine
from contextlib import asynccontextmanager

analyzer = AnalyzerEngine()

@asynccontextmanager
async def check_text(text: str):
  pii_results = analyzer.analyze(text=text, language="en")

  if pii_results:
    response = await cl.AskActionMessage(
      content="PII detected",
      actions=[
        cl.Action(name="continue", payload={"value": "continue"}, label="✅ Continue"),
        cl.Action(name="cancel", payload={"value": "continue"}, label="❌ Cancel"),
      ],
    ).send()

    if response is None or response.get("payload").get("value") == "cancel":
      raise InterruptedError

  yield

# ...

@cl.on_message
async def main(message: cl.Message):
  async with check_text(message.content):
    # This block is only executed when the user press "Continue"
    response = await cl.Message(
        content=f"Received: {message.content}",
    ).send()
If your application has a requirement to anonymize PII, Presidio can also do that. Modify the check_text context manager to return anonymized text when PII is detected.
Code Example
from presidio_anonymizer import AnonymizerEngine

anonymizer = AnonymizerEngine()

@asynccontextmanager
async def check_text(text: str):
  pii_results = analyzer.analyze(text=text, language="en")

  if pii_results:
    response = await cl.AskActionMessage(
      content="PII detected",
      actions=[
        cl.Action(name="continue", payload={"value": "continue"}, label="✅ Continue"),
        cl.Action(name="cancel", payload={"value": "continue"}, label="❌ Cancel"),
      ],
    ).send()

    if response is None or response.get("payload").get("value") == "cancel":
      raise InterruptedError

    yield anonymizer.anonymize(
      text=text,
      analyzer_results=pii_results,
    ).text
  else:
    yield text

# ...

@cl.on_message
async def main(message: cl.Message):
  async with check_text(message.content) as anonymized_message:
    response = await llm_chain.arun(
      anonymized_message
      callbacks=[cl.AsyncLangchainCallbackHandler()]
    )