What's New
The record of what we've built. Releases, improvements and milestones, newest first.
Rideau AI has run on hardware we own and operate in Canada since July 24, 2026, and everything we build now stands on that foundation. This page shows where we are and how we got here.
v1.2.0 Documents (November 16, 2026)
The first release built on our finished move to hardware we own. Ask questions across several documents and open the cited page beside the answer. Underneath, we rebuilt how the app handles files, because almost everything else depends on them. An uploaded file now goes into a virtual file system scoped to your conversation, and the model gets tools to navigate, search, read and write it. Modern models are trained to work with files and tools, so this plays to their strengths. Each model stays contained to the files in that conversation, with no access to a real file system.
Documents
- Document reading and indexing run on our hardware when you convert, preview or send a file. The processed content is reused in the conversation, so the AI can search and read relevant passages without you uploading the document again.
- The document viewer shows you the real pages, not just the extracted text.
- Canvas is now a document under Files, migrated automatically.
Redact
- Redact now works on documents as a whole. The same name, address or ID gets a consistent placeholder across the files in a conversation.
- Redact remembers everyone it has masked in a conversation, so a name found anywhere stays masked everywhere: earlier messages, files and later turns.
- The redaction preview shows exactly what will be sent, file names included. Cards and documents the AI creates show you the real names instead of placeholders.
- A final check runs on every Redact request just before it leaves for an outside model. If a masked name slipped through, the request is masked again or stopped.
- Redact catches names with or without accents, more Canadian health card and driver's licence formats, and dates of birth written in English or French. It also masks fewer everyday words by mistake.
- Countries, provinces and courts stay visible so the model can tell which law applies. Addresses are still masked.
- Web searches in Redact never name someone the conversation has masked.
- In Redact and Private, links to shared files (Google Drive, Dropbox, OneDrive and the like) aren't opened. Download the file and attach it instead.
- Redact now stops and asks you to try again when its scanner cannot finish in time, instead of sending a message with only the partial redaction it had.
- Added Try Redact: test redaction on your own text before you sign up. Nothing you paste there is stored.
Models
- A new model picker. Recommended shows the choices for the mode you're in (Quick, Everyday and Deep in Direct and Redact; Everyday in Private), and All models lists everything. Every model is labelled with who runs it, where it's processed and what the provider keeps (how we label models).
- A new model lineup from OpenAI, Anthropic and Google. A conversation on a model with a newer version moves to it automatically. A conversation on a model with no successor stays readable, and you can start a new chat to continue.
- GPT models now run on OpenAI's own service rather than Microsoft Azure. Only our AI processes your requests in Canada; every other model's label shows where it can be processed.
- Model labels now say when a provider runs on several cloud companies, and where those companies are from.
- Each answer now shows which model wrote it, including in conversations whose model has been retired.
- Model nicknames like The Workhorse are retired. Models appear under their own names, with operator, processing location and retention details. Existing conversations keep working.
- "The Cadet" model is retired. A conversation that used it still opens and carries on with Gemma, our private model, and each answer shows which model wrote it.
- Tasks are no longer in the picker. Conversations you started from a Task keep working.
Private
- Private adds a thinking option for harder questions. It's slower, and more likely to say it doesn't know than to guess.
- Conversation titles are now written by our AI on our hardware, whichever model you use.
Chat and capacity
- A dropped connection no longer loses an answer; the stream reconnects and carries on. Answers can run to 32,768 tokens on every plan, and we tell you if one reaches that ceiling.
- French conversations now run on French instructions rather than English ones to help keep some model responses in French.
- Capacity multipliers now use a fixed baseline: 1× is about one typical message on our private model. The numbers come from measured workloads and count the thinking a model does.
- Gemini turns use more of your capacity than before, because we now count the thinking a Gemini model does before it answers. Monthly web-search limits now apply in every month of a subscription. Nothing is charged retroactively in either case.
Account and billing
- Founding Member pricing is now closed to new subscribers. Thank you to everyone who joined us early. If you're a Founding Member, you keep paying the beta price for your plan while we're in beta. When we leave beta, your 20% Founding Member discount is applied to the new published price, for as long as your subscription stays continuous (no lapse longer than 30 days). Terms, section 15.6
- Signing up is shorter. Your recovery code is on one page, with an optional recovery email, and confirming your email in the browser you paid from carries straight on.
- Confirming a sign-in, a sign-up or an account recovery from an email, in the same browser, carries on in that tab. The other tab says so instead of signing you in twice.
- Regenerating your recovery code, or adding or removing a recovery email, works right after you confirm with your authenticator.
- Changing your name or location in Settings applies to new conversations, and conversations already open are told about the change.
- If your bank asks you to confirm a payment when you change plans or add seats, you confirm it in the app.
- If the 7-day refund isn't available, Billing tells you why.
- Plans have French names on French pages: Démarrage, Pro, Équipe Pro.
Notices
- Updated Terms of Service and Privacy Policy. Everyone accepts them once at their next sign-in, including confirming they are 18 or older.
- Only your own messages and files count toward an account warning. When our safety check blocks one of the AI's replies, it no longer counts against you.
- Product-update emails no longer track opens. Account emails, such as sign-in links and receipts, still tell us when they are opened, so we can help if one seems to be missing.
- Everyone was signed out once when this update landed, because we cycled session keys. Nothing was lost.
v1.1.2 Canadian Hardware Phase 3 (July 24, 2026)
- Our own web and database servers are deployed in Ontario.
- The app and its database left Azure for hardware we own and operate in Ontario. Your conversations, your files and your account now live on hardware that is ours.
- Protected in-app payloads have a second encryption layer inside TLS: the browser encrypts them and our application decrypts them. Public pages and sign-in flows have different paths. Cloudflare still handles connection metadata. See the encryption explanation.
- The app has not visibly changed, but the foundation has. We now run the web app, storage, safety scanning, document processing and our AI on hardware we own in Ontario. In Private, no outside AI provider receives your messages or files. External services, including link retrieval, have separate boundaries described on Where your data goes.
v1.1.1 Foreign document processors removed (June 6, 2026)
- The cloud-based document conversion service is decommissioned. Every document (PDF and text) is now processed exclusively on hardware we own in Canada.
- File uploads are limited to PDF and text until a Canadian-hosted conversion service can handle other formats. Convert other files to text, or print them to PDF, before uploading.
v1.1.0 Canadian Hardware Phase 2 (June 5, 2026)
- Our own GPUs are deployed in Ontario.
- Safety scanning, redaction and the models we host now run on hardware we own in Canada.
- PDFs are now processed entirely on hardware we own, instead of by a cloud-based document conversion service.
- Mistral (France) joins the model catalogue.
- PII Shield is renamed Redact.
- Multiple files can now be dropped into a conversation at once.
- If the context window can't fit your files, the app now stops before sending and explains why.
v1.0.5 Security hardening (May 16, 2026)
- Improved security for recovery codes, email changes, session lifetime, and providing a sign-out that takes effect everywhere at once.
v1.0.4 Launch week fixes (May 9, 2026)
- Fixes from launch week across billing, sign-in and team invitations.
- Thank you to our first paying customers, who arrived just hours after our launch announcement!
v1.0.0 Public beta launch (May 7, 2026)
- Public beta launch. Starter, Pro, Pro 2x and Team Pro plans.
- Full support for English and French from day one.
v0.3.0 Canadian Hardware Phase 1 (May 3, 2026)
- Our first servers, owned and run by us in Canada, are deployed. They hold the master encryption keys, outside the cloud provider.
- The cloud provider no longer holds the master keys used to encrypt stored conversation content.
- Keeping the keys on hardware we own separates their custody from the cloud storage. It is a protection around stored content, not immunity from lawful requests.
v0.2.0 First deployment (March 27, 2026)
- Rideau AI is live in Azure's Canada Central region, with every conversation encrypted by our application before it reaches the database. The hardware is rented from a foreign company; the encryption is ours.
v0.1.0 PII Redaction (February 26, 2026)
The idea behind Redact: use an outside model while sharing fewer identifying details. For a document such as an expense report, placeholders let the model work with a redacted version. Redaction is a tool you choose, not a guarantee that every identifying detail is removed or that using AI meets every professional obligation.
- Content safety and personal-information detection shipped, under the name PII Shield. Now called Redact, it replaces identifying details before an outside model receives a request when you choose that mode.