What Ask AI sends (and doesn't)

A precise answer for the security review — exactly what data leaves your browser when you use the Ask AI assistant, and what it never sees.

Before rolling this out to a team, the governance question is fair: what actually happens when someone asks the assistant a question on a page that might be open next to real tenant config? Here's the exact answer, not a marketing gloss.

What gets sent

When you ask a question — on a docs page, or via the sparkle button on a code block — three things leave your browser and go to Anthropic's API to generate an answer:

Your question text.
The current page's URL (e.g. /docs/onelake/v-order), sent as context so the assistant knows what you're reading.
fabdocs.dev's own documentation content — the assistant runs a search over this site's public docs and includes the matching pages so it can cite them, rather than guessing.

That's the complete list. See the request body in src/app/api/chat/route.ts if you want to verify this against the source directly — the site's code is public.

What never gets sent

Nothing about your Fabric tenant, workspace, or capacity. The assistant has no connection to your Microsoft Fabric environment, no API access, and no way to query it. It only knows what's published on fabdocs.dev.
No account or identity data. The free tier requires no sign-up; the Pro tier's subscriber check happens separately (see Privacy Policy) and isn't part of the message sent to Anthropic.
Whatever you don't type. The assistant can't see your screen, your other browser tabs, or anything outside the text box you typed into and the page URL you were on.

What fabdocs.dev itself retains

Nothing, beyond a per-IP request count used for the daily rate limit (resets every 24 hours — see Privacy Policy). Question text and answers are streamed directly to your browser and not logged or stored server-side.

The one thing to actually be careful of

Don't paste real secrets, connection strings, tenant IDs, or customer data into a question. The assistant doesn't need them to answer a question about Fabric concepts, and once submitted they've left your browser and are subject to Anthropic's own data handling terms — the same as pasting them into any third-party AI tool.

A safe pattern: ask "how do I write an idempotent MERGE" rather than "here's my actual notebook with real customer table names, what's wrong with it."

For the compliance record

Question	Answer
Does the assistant have access to our tenant?	No — no connection exists
Is chat history logged by fabdocs.dev?	No — streamed only, not stored
What third party processes the data?	Anthropic (the model provider)
Is a DPA available?	Not at this stage — this is a self-serve tool, not an enterprise contract. If your organisation needs one before allowing use, treat it the same as any other public AI chat tool for now
Can it be disabled for a team?	There's no admin control — it's a public, per-visitor tool. If that's a blocker, the free docs remain fully usable without ever opening the assistant

This page and the Privacy Policy are the two sources of truth here; if anything in this manual's marketing copy elsewhere ever seems to overstate what the assistant does, this page wins.