Masking PII/PIFI Data in LangFuse Traces

[guptadna]

Overview

We need guidance on implementing masking for PII (Personally Identifiable Information) and PIFI (Personal Information Financial Information) in LangFuse traces. LangFuse doc talks about masking data before sending data it to LLM and unmasking after that. While we process these types of data through LLMs, we want to ensure proper masking before the traces are sent to LangFuse.

Questions

1. What are the recommended approaches for masking PII/PIFI data before it reaches LangFuse traces?
2. Are there built-in masking features in LangFuse we should be using?
3. If implementing custom masking, at what point in the trace pipeline should this occur?

[inkeep-ai-assistant[bot]]

1. Recommended Approaches for Masking PII/PIFI Data
   Langfuse supports masking sensitive data by allowing you to define a custom masking function. This function processes all event inputs and outputs before data is sent to Langfuse. You can use regular expressions for redacting patterns such as emails, phone numbers, or credit card numbers, or leverage external libraries like llm-guard for more advanced anonymization(1)(2)(3).

2. Built-in Masking Features
   Langfuse provides built-in support for masking via the mask parameter on the client. This feature is available in the Python and JS/TS SDKs. When you initialize the Langfuse client, you can pass a masking function that will be automatically applied to all trace data (inputs, outputs, metadata) before it is transmitted to Langfuse(1)(2).

3. Where to Apply Custom Masking
   The masking function should be provided when initializing the Langfuse client. This ensures all relevant data is processed through your masking logic before being sent to the Langfuse backend. In the Python SDK, you do this by passing your function as the mask argument to the Langfuse constructor. The masking occurs on every event input/output as it is traced, so sensitive data never leaves your environment unmasked(1)(2).

Example (Python SDK):

import re
from langfuse import Langfuse

def masking_function(data, **kwargs):
    # Mask email addresses
    data = re.sub(r'\b[\w.-]+?@\w+?\.\w+?\b', '[REDACTED EMAIL]', data)
    # Mask phone numbers
    data = re.sub(r'\b\d{3}[-. ]?\d{3}[-. ]?\d{4}\b', '[REDACTED PHONE]', data)
    return data

langfuse = Langfuse(mask=masking_function)

(1), (2)

For more advanced needs (like using llm-guard or other PII detection libraries), you can integrate those inside your masking function(1)(3).

📚 Sources:

1. Masking of Sensitive LLM Data
2. Advanced Usage
3. LLM Security & Guardrails

---

Have another question? Just tag @inkeep.

[guptadna]

Thank you!

[clemra]

Glad it was helpful :)

[umudhasanli]

For the JavaScript and TypeScript SDK, Flare Redact can be used inside the Langfuse mask callback. It redacts nested objects locally, validates formats such as cards and national IDs, and never needs to expose the matched value in logs. The best place is the client boundary before the trace enters the send queue.

I maintain it here: https://github.com/umudhasanli/flare-redact