1. What Nexcerpts does with your data
Nexcerpts connects to the systems your company already runs, indexes their contents, and answers questions in plain language with links back to the source.
To do that it must read your content. It does not need write access to anything, and every connector requests read-only scopes unless a specific feature — drafting a reply, creating a ticket — requires otherwise and you enable it explicitly.
We do not train any model on your data. Your content is never used to train, fine-tune, or improve a model, ours or a third party's.
2. The permission model
This is the part most enterprise search tools get wrong, so it's first.
Indexing is not access. Nexcerpts reads everything it's connected to, but it also copies who is allowed to see each item, from the source system, per document and per person.
Mirrored per item. Permissions come from the source system's own ACL for that specific document, not from a workspace-level or folder-level approximation.
Checked at query time. Every answer re-evaluates the asking user's access at the moment they ask. Permissions revoked in the source take effect on the next question, not the next crawl.
Fails closed. If we cannot determine who is allowed to see an item — the source ACL is missing, stale, or the connector cannot retrieve it — that item stays private to the person who connected the integration. It is never widened to your whole workspace to make search look better.
No existence leakage. A user who lacks access sees no banner, no greyed-out row, no “1 result hidden.” Telling someone a document exists is itself disclosure.
The practical consequence: three people asking the same question get three different answers, and none of them can infer what the others saw.
3. Where your data lives
Managed cloud (default). Runs in our infrastructure on AWS. Your content is stored in a Postgres database with vector embeddings, plus object storage for files. Disks and snapshots are encrypted at rest.
Bring your own cloud (BYOC). The entire stack deploys inside your cloud account, and an air-gapped mode enforces that model and embedding endpoints stay inside your network. Your content and the inference over it never leave your boundary. This is the option for teams who cannot place company content with a third-party processor.
4. Subprocessors
Under the managed cloud model, the following third parties process customer content. Under BYOC you choose which of these apply and can substitute your own model endpoints.
| Processor | Purpose | Data sent |
|---|---|---|
| Anthropic | Answer generation | Retrieved passages relevant to the question |
| OpenAI | Text embeddings | Document text, at index time |
| AWS | Hosting, database, object storage | All indexed content |
| Deepgram | Speech-to-text | Meeting audio, if meeting capture is enabled |
| Recall.ai | Meeting recording | Meeting audio/video, if meeting capture is enabled |
| Tavily | Web search | The query only, never your content |
| WorkOS / BoxyHQ | SSO, SAML, SCIM | Identity and directory metadata only, never content |
Read this plainly: in the managed model, passages of your content are sent to Anthropic and OpenAI to generate answers and embeddings. Neither trains on it. If that transfer is unacceptable to your policy, BYOC is the answer, and it's why we build it.
Meeting capture (Deepgram, Recall.ai) is off unless you turn it on.
5. Encryption
In transit: TLS 1.2+ on every connection, including to subprocessors.
At rest: database and object storage encrypted at the platform level, AES-256.
Connector credentials: encrypted with AES-256-GCM before they touch the database, so a database compromise alone does not yield usable tokens to your connected systems.
Key separation: connector credentials are encrypted under a dedicated key, held separately from the one that signs user sessions. The two rotate independently, so responding to a compromised session key never means dropping your live integrations.
6. Authentication and access control
SSO via SAML 2.0 and OIDC.
SCIM provisioning — deprovisioning in your IdP removes access here.
Role-based access with four workspace roles: Super Admin, Admin, Setup Admin, Member.
Invite-only by default — no self-registration against your workspace domain.
Admin console for connector allowlisting, so an admin controls which systems may be connected at all.
Permission audit trail — grants, revocations, denied access attempts, share links created and revoked, and administrative cross-user reads are all recorded.
7. Retention and deletion
Source of truth stays yours. Deleting a document in Slack, Drive, or Confluence removes it from Nexcerpts on the next sync; soft-deleted content is purged by a scheduled retention job after 30 days.
Disconnecting a connector removes its indexed content.
Account deletion removes your workspace data. Available on request; we will confirm in writing when it completes.
8. What we do not have yet
We would rather you hear this from us than find it in diligence.
SOC 2 Type II — not certified. Our architecture is built to the controls and we can walk you through them, but we do not hold the report. If your policy requires it before signature, we are not the right vendor this quarter. Tell us and we'll say so plainly.
Penetration test — no third-party report yet.
ISO 27001 / HIPAA / FedRAMP — not held.
Read-access audit — we record permission changes and search activity, but not a per-document view trail. If your policy requires one, ask and we'll tell you where we are with it.
Customer-configurable retention — retention is currently a fixed 30-day policy, not a setting you control.
Uptime SLA — contractual SLA available; no published historical figures, because we are early and would be inventing them.
We are an early company. The controls above are real and you can verify them. The certificates are not there yet, and we won't imply otherwise.
9. Questions this document doesn't answer
Send them to [email protected] and you'll get a direct answer, including “we don't do that yet” where that's the truth. If you have a standard questionnaire — CAIQ, SIG, your own — send it and we'll complete it rather than send you a link to a trust page.