Private by architecture, so there is less to take on trust
AiGrid has no cloud service and no vendor account. The documents, the extracted text, the index, and the models all stay in your environment. This page sets out what that rests on, and what the beta does not yet cover.
Where your data goes
- Nothing leaves your network
- Document processing and search call no external service. All models ship inside the product images, so nothing is downloaded while it runs.
- Read-only access to your documents
- The appliance is granted read access to your document store and never writes to it. Your originals remain the source of truth, and the index can always be rebuilt from them.
- Air-gapped capable
- Once installed, AiGrid needs no internet access. The documentation covers how to carry out the installation in an environment without internet access.
- Encrypted in transit
- All traffic to the appliance is HTTPS from first boot, using a private certificate authority that the appliance creates at setup.
What a borrowed machine can and cannot do
Lending a desktop to a processing pool is only acceptable if a compromised desktop gains nothing from it. The worker is built around that.
- No inbound connections
- Every call is outbound HTTPS, from the machine to the appliance and to the document store. No port is opened on any desktop.
- Transient storage only
- The agent uses temporary storage while it converts a document. The originals stay in your document store.
- Its own service account
- On a desktop the agent runs under a specific service account, not the user's account, and at low priority so it does not compete with other work.
- A schedule you control
- An administrator sets the schedule for each pool. It can be out of hours only, or around the clock on unused compute such as a DR site.
Retrieval without generation
AiGrid returns passages from your documents, each with its filename and page number. No model in the product writes answers or summaries, so the product itself cannot invent one.
What your assistant does with those passages is the assistant's behavior. AiGrid gives it real sources to cite, and flags results from documents that could not be read well so they are not treated as evidence.
Access and supply chain
- Operators sign in through your own identity provider using OpenID Connect, with a local administrator account as a fallback.
- The retrieval endpoint is HTTPS only and requires a search token.
- Every software dependency is pinned to an exact version and hash.
- The whole stack, including every model, is permissively licensed (Apache 2.0, MIT, BSD). Copyleft and non-commercial components were excluded.
What the beta does not cover yet
You will find these in a security review, so here they are up front.
- Per-user result filtering
- One search token reads everything indexed. Results are not filtered by the permissions of the person asking, so index only what every user of the endpoint is allowed to see.
- Dashboard roles
- Everyone who can sign in to the dashboard is an administrator.
- Encryption at rest
- The appliance does not encrypt its own storage. Use disk-level encryption on the host.
- Certifications
- We do not claim any compliance certification for AiGrid. Keeping data in your own environment supports data residency and sovereignty requirements.