Written for the person who has to say yes

Architecture: deliberately small

Each client runs a dedicated, single-tenant deployment: one service, one background worker, one PostgreSQL database. There is no shared infrastructure between clients, no multi-tenant data partitioning to trust, and no cross-tenant attack surface — a new client is a new deployment.

Everything stateful lives in that one database: document chunks and their access-control metadata, conversations, the audit log, background jobs and evaluation results. That gives you one backup scope, one encryption scope, and one data map — the database schema itself. When your DPO asks “where exactly is our data?”, the answer is one sentence.

Access control below the AI

Every document carries a list of groups permitted to read it. When someone asks a question, retrieval filters on those permissions inside the database query — restricted documents are excluded before the language model is ever invoked.

• The model cannot leak what it never receives. No prompt injection, jailbreak or phrasing trick can cross a group boundary, because the boundary is enforced in SQL, not in instructions to the model.
• All surfaces — web chat, API, CLI — sit above the same core. Access control, audit and refusal behaviour live below the surface layer and cannot be bypassed by any of them.
• Permission changes take effect immediately and are tracked separately from content, so changing who may read a document never requires re-processing it.

Groundedness: cite or refuse

Naxis answers only from retrieved passages, and every answer carries citations to them. When no permitted document supports an answer, it refuses. This is a structural rule — not a confidence threshold that can drift or need calibration. The refusal behaviour is itself measured: built-in evaluation includes questions the assistant should decline, and reports how often it correctly does.

Data flow and models

Document embeddings are always computed inside your deployment — in every edition, your documents are never sent to a third party to be indexed. The editions differ only in where answer generation runs:

Source credentials (service-account keys, API tokens) are write-only once saved: they can be replaced, but never read back through the interface or API. Every connector is designed to work with read-only, minimally-scoped credentials.

Tamper-evident audit log

Every authentication, question, retrieval and administrative action is appended to an audit log in which each entry includes a hash of the entry before it. Altering or deleting any record breaks the chain, and a built-in verification recomputes the entire chain on demand — from the admin interface or the command line.

The log is pseudonymous by design: it records opaque actor identifiers and chunk references, not names or document contents, so the audit trail itself stays lean under data-protection review.

GDPR, in the product

Compliance tooling ships inside Naxis rather than in documentation:

• Records of processing (Art. 30). A processing manifest — data categories, retention, models, sub-processors, technical measures — is generated from the live deployment configuration. It cannot drift from reality, because it is never hand-written.
• Subject access (Art. 15). Export everything held about a person with one command or click. Users can also export their own data themselves, without asking anyone.
• Erasure (Art. 17). Erase a person’s conversations, or a document’s data, with built-in auditable tooling. Deleting a user account can include full erasure of their history.
• Retention. A nightly job enforces the retention schedule automatically:

Conversations are stored keyed to opaque subject identifiers, never names. The knowledge-base contents are yours: what personal data they contain is under your control, and the corpus simply mirrors your sources.

EU AI Act

Naxis Assistant is built to the transparency obligations of Article 50: every surface discloses that answers are AI-generated, every answer is stamped with the model that produced it, and every answer carries source citations. High-risk uses under Annex III are contractually excluded — Naxis is a knowledge-access tool, and its terms keep it one.

Security questions and reviews

We’re glad to complete security questionnaires, walk your team through the architecture, and provide the generated compliance manifest for a trial deployment so you can review the real configuration rather than a sample. For any of these — or to report a security concern — contact us or write to hello@naxis.tech.