Mask / redact PII from traces logged to langfuse

[marcklingen]

Describe the feature or potential improvement

In todays community session, the idea came up multiple times to redact PII and other sensitive information from traces.

I'd suggest to add a callback on the integration level which can be optionally used to remove sensitive data before it is logged to Langfuse.

Additional information

Potential DX (low-level sdk, extends to other integrations)

import re

langfuse = Langfuse(
    # redact emails
    mask=lambda input: re.sub(r'[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}', '[REDACTED]', input)
)

Probably a util is necessary here to apply the mask to more complex (size/objects) inputs/outputs without performance issues.

[ajram23]

@marcklingen Do you folks plans to use a library like https://microsoft.github.io/presidio/ cuz phone number formats/emails/etc can be pretty crazy!

[marcklingen]

The idea here is that Langfuse could support a masking callback which then let's you use any PII masking library depending on your use case and goals. Open issue for this: #3518

[ajram23]

ok will wait to see what you folks cook up before chiming in.

[chai3]

#1676
Can it also be used to reduce huge input/output?

If so, filter is better.

langfuse = Langfuse(
    filter_input=lambda input: input[0:1000]),
    filter_output=lambda output: output[0:1000]).
)

langfuse = Langfuse(
    filter=lambda data, type: input[0:1000]) # type is "input" or "output"
)

[hassiebp]

Passing a custom masking function has been released, please check out our docs, try it out and let us know your feedback! 🙏🏾

• Docs
• Python SDK release
• JS SDK release