PII never reaches the LLM. Zero retention is a toggle away.

Every question is scanned for PII (emails, names, phone numbers, card numbers, IDs) and masked before it reaches any model. Workspaces can go further and enable zero retention — questions and answers are streamed and dropped, never persisted. Both are on by default where they can be, and controllable per workspace where they can't.

Start FreeSecurity Overview

What PII masking + zero retention actually do

Two controls, deployed correctly — and audit-friendly

Before the LLM, before storage

PII Masking

  • Detects emails, phones, credit cards, common ID formats, names
  • Applied to every question before any LLM call — including reranking + eval calls
  • Applied to every stored message — the DB holds the masked form
  • Language-aware detectors, extendable per workspace
  • Cited answers still reference the underlying doc, not the raw user question
Off by default, on with one setting

Zero Retention

  • Per-workspace toggle: questions + answers stream but never persist
  • No conversation row is created; no feedback can be attached
  • Still generates cited, grounded answers in real time
  • Compatible with SSO, restricted documents, and per-KB permissions
  • Exempt from the monthly-quota counter, since nothing is stored to count
Built for the security review

Compliance Posture

  • Workspace-scoped data isolation across content, embeddings, and chats
  • Encryption in transit and at rest
  • Your data is never used to train models — none of it, ever
  • Data-processing agreement (DPA) available on Enterprise
  • Regional data residency (US / EU) on Enterprise
Why this exists

The two questions every security review asks

'Does user PII reach a third-party model?' and 'Where does the conversation get stored?' are the two questions that stall AI procurement. PII masking answers the first before it can be asked — nothing sensitive ever leaves our tenant to reach the LLM. Zero retention answers the second — for workspaces that choose it, the answer is 'nowhere; it streams and it's gone'. Both are architectural controls, not aspirational policy.

  • Masking runs before every LLM call: chat, rerank, eval, deflect, agent
  • Zero-retention workspaces skip persistence entirely — no message row exists
  • Audit logs cover admin actions; traces show retrieved chunks per answer
  • SSO (OIDC + SAML) and restricted docs stack on top
  • Enterprise adds DPA, regional residency, and dedicated support
What teams say

Deployed in production, cited by the buyers who chose it

“Deployed on our docs site in an afternoon. Every answer has citations, and the abstention gate means we've never had a customer complain about a made-up answer.”
PN
Priya Nair
Head of Customer Support · Supabase
“The knowledge base connected to our Slack, Confluence, and helpdesk in one setup. On-call teams get the same cited answer whether they ask in chat, in the widget, or from Cursor.”
TR
Tom Richter
IT Operations Manager · Grafana Labs
“The gap analytics turned into a real docs backlog. Deflection went up because we finally knew which pages were missing — the AI told us.”
AC
Ana Castillo
VP of Customer Experience · Clerk

Frequently Asked Questions

Common questions about PII & Zero Retention

Emails, phone numbers, credit-card numbers, common government-ID formats, and personal names — with regional and language-aware detectors. Enterprises can extend the detector list per workspace for domain-specific identifiers.
Before any LLM call and before storage. That includes the chat model, the reranker, the eval model, form deflection, and agent runs — every code path that would send text to a model sees the masked form, not the raw one. The database only ever holds the masked message.
Questions and answers stream to the caller in real time and are then dropped — no conversation row is created, no message is stored, no feedback can be attached. Retrieval still happens against your indexed content; the model still generates a cited answer. Only the transcript is not persisted.
Zero-retention workspaces are exempt from the monthly AI-question quota counter, because the counter counts stored user messages and there are none. Enterprise contracts price this differently; see pricing for details.
No. Indexed content, embeddings, chats, and feedback stay isolated to your workspace and are never shared across customers or used to train any model — ours or a third party's. This is contractual on paid plans and technical on every plan.
The providers BeforeQuery uses run under zero-retention API terms (no training on API inputs) as documented in their enterprise agreements. On Enterprise you can bring your own model credentials (BYOK) so the traffic runs on your own account, under your own terms.

Bring this to your security team

Read the security overview, or talk to us about DPA, regional residency, and BYOK on Enterprise.

Get Started Free