1 of 1

LangChain Prompt Sanitization Tutorial

LangChain Tutorial: Integrating Nightfall for Secure Prompt Sanitization

Generative AI systems like OpenAI's ChatGPT have revolutionized how we interact with technology, but they come with a significant risk: the inadvertent exposure of sensitive information (OWASP LLM06). Without proper safeguards, these AI platforms may receive, process, and potentially retain confidential data, including:

Personally Identifiable Information (PII)
Protected Health Information (PHI)
Financial details (e.g., credit card numbers, bank account information)
Intellectual property

Real-world scenarios highlight the urgency of this issue:

Support Chatbots: Imagine a customer service AI powered by OpenAI. Users, in their quest for help, might unknowingly share credit card numbers or Social Security information. Without content filtering, this sensitive data could be transmitted to OpenAI and logged in your support system.
Healthcare Applications: Consider an AI-moderated health app that processes patient and doctor communications. These exchanges may contain protected health information (PHI), which, if not filtered, could be unnecessarily exposed to the AI system.

Content filtering is a crucial safeguard, removing sensitive data before it reaches the AI system. This ensures that only necessary, non-sensitive information is used for content generation, effectively preventing the spread of confidential data to AI platforms.

Python Example

Let's examine this in a Python example using the LangChain, Anthropic, and Nightfall Python SDKs. You can download this sample code here.


import os
from dotenv import load_dotenv
from nightfall import Confidence, DetectionRule, Detector, RedactionConfig, MaskConfig, Nightfall
from typing import Dict, List
from langchain.chains.base import Chain
from langchain.schema.language_model import BaseLanguageModel
from langchain.schema.prompt_template import BasePromptTemplate
from langchain.prompts import PromptTemplate
from langchain_anthropic import ChatAnthropic
from langchain.schema.runnable import RunnableSequence, RunnablePassthrough
from pydantic import Field

# Load environment variables
load_dotenv()

# 1) Setup Nightfall
# By default Nightfall will read the NIGHTFALL_API_KEY environment variable
nightfall = Nightfall()

# 2) Define a Nightfall detection rule
detection_rule = [DetectionRule(
    [Detector(
        min_confidence=Confidence.VERY_LIKELY,
        nightfall_detector="CREDIT_CARD_NUMBER",
        display_name="Credit Card Number",
        redaction_config=RedactionConfig(
            remove_finding=False,
            mask_config=MaskConfig(
                masking_char="X",
                num_chars_to_leave_unmasked=4,
                mask_right_to_left=True,
                chars_to_ignore=["-"])
        )
    )]
)]

# 3) Classify, Redact, Filter Your User Input

# Setup Nightfall Chain element
class NightfallSanitizationChain(Chain):
    input_key: str = "input"
    output_key: str = "sanitized_input"

    @property
    def input_keys(self) -> List[str]:
        return [self.input_key]

    @property
    def output_keys(self) -> List[str]:
        return [self.output_key]

    def _call(self, inputs: Dict[str, str]) -> Dict[str, str]:
        text = inputs[self.input_key]
        payload = [text]
        try:
            findings, redacted_payload = nightfall.scan_text(
                payload,
                detection_rules=detection_rule
            )
            sanitized_text = redacted_payload[0] if redacted_payload[0] else text
            print(f"\nsanitized input:\n {sanitized_text}")
        except Exception as e:
            print(f"Error in sanitizing input: {e}")
            sanitized_text = text
        return {self.output_key: sanitized_text}

# Initialize the Anthropic LLM
llm = ChatAnthropic(model="claude-2.1")

# Create a prompt template
template = "The customer said: '{customer_input}' How should I respond to the customer?"
prompt = PromptTemplate(template=template, input_variables=["customer_input"])

# Create the sanitization chain
sanitization_chain = NightfallSanitizationChain()

# Create the full chain using RunnableSequence
full_chain = (
    RunnablePassthrough() |
    sanitization_chain |
    (lambda x: {"customer_input": x["sanitized_input"]}) |
    prompt |
    llm
)

# Use the combined chain
customer_input = "My credit card number is 4916-6734-7572-5015, and the card is getting declined."
print(f"\ncustomer input:\n {customer_input}")
try:
    response = full_chain.invoke({"input": customer_input})
    print("\model reponse:\n", response.content)
except Exception as e:
    print("An error occurred:", e)

Step 1: Setup Nightfall

If you don't yet have a Nightfall account, sign up here.
Create a Nightfall key. Here are the instructions.

Install the necessary packages using the command line:

pip install langchain anthropic nightfall python-dotenv

Set up environment variables. Create a .env file in your project directory:

ANTHROPIC_API_KEY=your_anthropic_api_key
NIGHTFALL_API_KEY=your_nightfall_api_key

Step 2: Configure Detection

Create an inline detection rule with the Nightfall API or SDK client, or use a pre-configured detection rule in the Nightfall account. In this example, we will do the former.

If you specify a redaction config, you can automatically get de-identified data back, including a reconstructed, redacted copy of your original payload. Learn more about redaction here.

Step 3: Classify, Redact, Filter Your User Input

to integrate content filtering into our LangChain pipeline seamlessly. We'll create a custom LangChain component for Nightfall sanitization. This allows us to seamlessly integrate content filtering into our LangChain pipeline.

Explanation

We start by importing necessary modules and loading environment variables.
We initialize the Nightfall client and define detection rules for credit card numbers.
The NightfallSanitizationChain class is a custom LangChain component that handles content sanitization using Nightfall.
We set up the Anthropic LLM and create a prompt template for customer service responses.
We create separate chains for sanitization and response generation, then combine them using SimpleSequentialChain.
The process_customer_input function provides an easy-to-use interface for our chain.

Error Handling and Logging

In a production environment, you might want to add more robust error handling and logging. For example:

import logging

logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

def sanitize_input(text):
    payload = [text]
    try:
        findings, redacted_payload = nightfall.scan_text(
            payload,
            detection_rules=[detection_rule]
        )
        if findings:
            logger.info(f"Sensitive information detected and redacted")
        return redacted_payload[0] if redacted_payload[0] else text
    except Exception as e:
        logger.error(f"Error in sanitizing input: {e}")
        # Depending on your use case, you might want to return the original text or an error message
        return text

Usage

To use this script, you can either run it directly or import the process_customer_input function in another script.

Running the Script Directly

Simply run the script:

python secure_langchain.py

This will process the example customer input and print the sanitized input and final response.

Using in Another Script

You can import the process_customer_input function in another script:

from secure_langchain import process_customer_input

customer_input = "My credit card 4916-6734-7572-5015 isn't working. Contact me at alice@example.com."
response = process_customer_input(customer_input)
print(response)

Expected Output

What does success look like?

If the example runs properly, you should expect to see an output demonstrating the sanitization process and the final response from Claude. Here's what the output might look like:

> Entering new SimpleSequentialChain chain...

> Finished chain.

Sanitized input: The customer said: 'My credit card number is XXXX-XXXX-XXXX-5015, and the card is getting declined.' How should I respond to the customer?

Final Response: I understand you're having trouble with your credit card (XXXX-XXXX-XXXX-5015) being declined. I apologize for the inconvenience. To assist you better, I'll need some additional information...