Data handling

"Content" means your prompts and chat messages, the CAD content read from the active document, the outputs generated for you, the screenshots the copilot captures, and the files you attach or submit as drawing standard references.

On Free, Student and Professional that content is retained so chat history and cross-session recall work, and may be used to operate, secure and improve the service, including to improve and train CADABRA’s own AI models. That is what the EULA and privacy policy state, and it is the default until you change it.

The providers whose models generate your output are contractually barred from training on your content and required to delete it after processing. That is a commitment from the provider, not a technical guarantee CADABRA can enforce.

On Enterprise the zero-retention arrangement is applied to your accounts as a configuration step when the agreement is set up. It does not switch itself on because of the plan tier. If you are on Enterprise and are not certain it was applied, ask for confirmation in writing.

The setting

The row is "Do not store my data" on the dashboard, under Preferences. It is off by default.

It is one account setting, not a per-machine one, so it follows you to every workstation you sign in on. The same switch appears in the task pane Settings under Data and privacy, where it reads "Don’t use my data to train AI". Changing either changes both.

When a plan pins it on, the row is disabled and says so rather than disappearing.

• It applies from the moment you turn it on. It is not retroactive, and it does not delete content already stored. For that, see deletion.
• You do not have to find it yourself. Email team@cadabrai.com and we will set it on your account.

What changes when it is on

It is enforced by the service, not by the panel, so it holds no matter which version of the add-in is installed. Four things behave differently, and three of them are features you lose.

Chat history is refused

Creating a chat returns an ephemeral id and retained: false. Appending a message or renaming a chat returns success and writes nothing. The panel shows a line on the history view saying the conversation is not retained.

The consequence: no chat history, no resuming a conversation on another machine, and no cross-session recall.

Indexing and retrieval return nothing

The index request returns indexed: 0, retained: false and stores no vector. Retrieval returns an empty result list, so the document matches described in what leaves the workstation are never fetched and never appended to a prompt. The copilot works from the live situation and the digest alone.

Attachments are not uploaded

A file you attach is passed to the model to answer that turn and is not written to storage, so it is not on the chat afterwards and there is nothing to fetch later. Limits are unchanged: see attachments.

A file you add to a drawing standard is not a chat attachment. It is uploaded and stored under this setting too: see what is kept anyway.

Quality traces keep no content

The observability trace for a turn replaces the prompt text, the output, the model’s reasoning and the per-operation text with [redacted: zero-retention], and your name and email are not attached to it. What is left is metadata: model identifier, mode, plan, latency, stop reason, token counts, and whether the step succeeded. Usage telemetry carrying your content is dropped rather than stored.

What is kept anyway

These categories survive the setting and survive an Enterprise zero-retention arrangement.

• Operational diagnostics. The logs CADABRA keeps about the service, and your machine name. They expire 90 days after they are written, and per-account erasure from them before that is not something we can carry out.
• Feedback you explicitly submit. A thumbs down sends what you typed, the request id and your version. Where your build includes them, that turn’s prompt and output go with it, because that is the substance of the report.
• Drawing standard reference files and the learned standard, which the feature exists to store for you. Removing a file from a standard deletes the stored copy; deleting a standard deletes every file under it.
• Usage records for billing, quota and abuse prevention. They record the model, the mode, token counts and the chat identifier, not the text of your prompts or outputs.
• De-identified, aggregated metrics that no longer identify you or your designs.

Your learned drawing standard and its reference files are reachable by authorized CADABRA staff under least-privilege controls, for support.

The log files CADABRA writes on your own machine are a different thing. They are yours, nothing in CADABRA deletes them, and they reach us only when you attach one: the files you can attach.

What leaves the workstation on a prompt

Three things, plus your prompt. All of it is derived from the open document. None of it is the document.

The live situation

Whether a document is open, and if so its name, type, full local path, active configuration, configuration count and units. Whether you are in a sketch, the feature count, and any rollback state. Your current selection. A summary of the assembly. Up to 12 features in error or warning.

The full local path is in the context of every turn, not only in diagnostics, and a Windows path carries your folder names and your user name.

The bounded digest

For a part or an assembly, a digest capped before it is sent: up to 60 features, up to 16 custom properties with each value trimmed to 120 characters, and a geometry block of bounding box, mass, volume, material and body count. The whole digest is capped at 3,000 characters.

The retrieval index

When you open a part, the panel builds text chunks and sends them to be embedded and stored: one document line of title, active configuration, component count and mass, then one chunk per feature-tree node up to 400 of them, skipping the History and Sensors folders. The document is keyed by a hash of its file path, not the path. On a later prompt the top 6 matches are fetched and appended.

It is skipped for assemblies, skipped for drawings, and skipped for a part over 1,500 features, because walking those trees on open freezes the SOLIDWORKS UI thread. It is refused outright by the service for an account with "Do not store my data" on, or in encrypted mode.

Screenshots

Produced by SOLIDWORKS’ own image export, so they contain the model view and nothing else. The panel does not read your screen: no desktop, no other application window, and no part of the CADABRA panel. If the export fails, no screenshot is taken and the turn continues without one.

Your CAD files

Your part, assembly and drawing files are not uploaded. The panel sends the derived content above.

There is one exception, and it is one you trigger on purpose. Drawing standard reference files are uploaded in full and stored, keyed to your organization, so the learned standard stays reproducible. If you submit a part or assembly alongside the drawing, both are uploaded. Files stay until you remove them; there is no automatic expiry.

A .slddrw attached to a chat outside Drawings Mode is different: it is opened on your machine, captured as an image, and only the image is transmitted. The drawing file itself is not uploaded.

Attach the same file in Drawings Mode and it is a drawing standard reference: it is uploaded in full and learned from, as above. See attachments in Drawings Mode.

Encryption

In transit, TLS on 443 to the hosts listed under install.

At rest, storage-level encryption across the databases and object storage, which cannot be disabled. CADABRA holds those keys, so access is limited by least-privilege controls rather than made impossible. An authorized member of staff can reach stored content.

A second layer sits above that one: stored chats, attachments, standards files and feedback are encrypted with AES-256-GCM before they are written, each account’s data key wrapped by a key held in a hardware-backed vault separate from the data, and decryption gated behind a key permission that storage access alone does not grant. Storage access without that permission reads ciphertext. It is one row in the task pane on every plan, "Encrypt my stored data at rest". While it is on, indexing and retrieval return nothing, the same as under zero retention.

Sign-in tokens held on your own machine are encrypted with DPAPI.

No at-rest measure changes the fact that answering a prompt means the model provider processes your content in plaintext.

Isolation and staff access

• Indexed content is filtered to the account that wrote it, and a retrieval request cannot reach another account’s chunks.
• Only the context a given prompt needs is loaded at run time. Your library is never uploaded wholesale.
• Staff access to stored content is granted through the cloud provider’s permission system, to the people who need it, for three reasons: investigating a failure you reported, keeping the service running, and responding to a security or legal obligation.
• The internal console shows account details, plan and subscription state, add-in version, devices and token totals. It cannot open your account, read your chats or open your files, and there is no impersonation path.
• Your designs, drawings and prompts are not used in demonstrations, sales material, screenshots, talks, case studies or public benchmarks. If we ever want to show your work, we ask first.

Single sign-on and role-based access are Enterprise features. See SSO and directory sync.

Deleting a chat, your content, or your account

A chat

Delete it in the panel. That removes the chat and its attachments.

Your account

Dashboard, Account page, Delete account in the danger zone. You confirm with your password, or without one on an account whose password is managed by your identity provider, and the button reads "Delete account permanently". It cancels the subscription, detaches payment history in Stripe and removes the account record and the identity with the authentication provider. Your account covers the page, billing covers what happens to the subscription.

Content stored before you changed the setting

Deleting the account does not by itself erase content already stored in the copilot backend, and closing it removes the identity used to locate that content. Email team@cadabrai.com before or at the same time as closing the account, and we will carry the erasure out and confirm when it is done.

Three limits apply.

1. Billing and tax records are kept up to 7 years, along with anything a legal obligation makes us hold. If we rely on one of those we say which.
2. One account’s records cannot be pulled out of operational telemetry, logs and diagnostics. Collection can be stopped for you; per-user erasure of that data cannot be promised today.
3. Databases are backed up continuously, so a deleted record can persist in backups for the length of the restore window.

Sub-processors

§6 of the privacy policy names every third party and states what each receives. In outline:

• Identity. Microsoft Entra External ID, and Amazon Cognito during the migration. Your password is stored and verified there, not by CADABRA.
• Model providers. Amazon Bedrock and Microsoft Azure AI Foundry. Which one answers depends on the deployment your request reaches.
• Observability. Langfuse, hosted in the United States and outside CADABRA’s cloud accounts, which holds the quality traces described above. Azure Monitor and Application Insights receive traces of the services themselves, not your content.
• Payments. Stripe. Card numbers stay with Stripe.

CADABRA runs as an add-in on your workstation and reaches a hosted processing environment for inference. It does not run offline.

The servers, databases and object storage are in the United States, as are the identity, payment and observability providers. Model deployments use a global routing tier, so the inference request that generates your output may be served by provider capacity in a data center outside the United States.

Reporting a vulnerability

Send it to team@cadabrai.com. The same address is published at /.well-known/security.txt. A product bug that is not a security issue goes through troubleshooting instead.