Loading video...

Video Failed to Load

Go Home

Today we're launching sensitive data redaction in Mastra observability! Automatically strip out your users’ PII from agent traces:

11,577 views • 2 months ago •via X (Twitter)

19 Comments

Sam Bhagwat's profile picture
Sam Bhagwat2 months ago

𝚂𝚎𝚗𝚜𝚒𝚝𝚒𝚟𝚎𝙳𝚊𝚝𝚊𝙵𝚒𝚕𝚝𝚎𝚛 is enabled by default and you can customize fields. Here's the blog post:

Gautam Paranjape's profile picture
Gautam Paranjape2 months ago

amazing! definitely really valuable

Sam Bhagwat's profile picture
Sam Bhagwat2 months ago

yeah a really useful implementation of our processor api I think

Bob Remeika's profile picture
Bob Remeika2 months ago

🚢🚢🚢

Benjamin Laval's profile picture
Benjamin Laval2 months ago

awesome

Sam Bhagwat's profile picture
Sam Bhagwat2 months ago

Thanks!

Mustafa Elsayed's profile picture
Mustafa Elsayed2 months ago

Hi @calcsam I am using mastra with Pg as a store and memory, and noticed a delay about 3 seconds when connecting with PG, Is there anything that might produce such effect so I can revisit and enhance. Thanks

Brennan McEachran 👨‍🚀's profile picture
Brennan McEachran 👨‍🚀2 months ago

Oft. i see PII redaction maintaining some traceability. Like, instead of "*" it's [EMAIL_1] Wonder how this will effect things like traces -> dataset -> evals

Sam Bhagwat's profile picture
Sam Bhagwat2 months ago

this is a good point. pinging Eric on our team who shipped this

Brennan McEachran 👨‍🚀's profile picture
Brennan McEachran 👨‍🚀2 months ago

I think my bias is coming from being exposed to this Which i think is a small model for this task

Sam Bhagwat's profile picture
Sam Bhagwat2 months ago

yeah eric and i talked about this and it makes a lot of sense to both of us. will be working on it

Brennan McEachran 👨‍🚀's profile picture
Brennan McEachran 👨‍🚀2 months ago

Appreciate you two

Dan Mullaney's profile picture
Dan Mullaney2 months ago

PII redaction in traces is the compliance win, but the same trace-level visibility is what most teams are still missing on the cost side. If you can strip a field out of every agent trace automatically, you can just as easily tag every trace with what it cost to produce. Is cost attribution on the observability roadmap too, or does Mastra expect teams to bolt that on separately?

Sam Bhagwat's profile picture
Sam Bhagwat2 months ago

we already have that

Dan Mullaney's profile picture
Dan Mullaney2 months ago

Good to hear it's already there. Is it opt-in per project or on by default, and does it break down to the step or tool-call level, or just the trace as a whole? That granularity is usually the difference between a nice dashboard and something people actually use to make decisions.

anirudh's profile picture
anirudh2 months ago

cool! didn't SensitiveDataFilter exist before too? what's new in this release 👀

Sam Bhagwat's profile picture
Sam Bhagwat2 months ago

yeah, but we never actually announced it lol

Boring Automations's profile picture
Boring Automations2 months ago

@mastra Yes having placeholders for redacted inputs is useful for maintaining context, for the agent receiving the user prompt and the downstream processes (eg. Observability, updating evals from real traces)

Praveen Yadav's profile picture
Praveen Yadav2 months ago

What we use for redaction?

Related Videos