AI Use Policy

What this covers

MindWiki ships several AI-powered features across iOS, macOS, and the MindWiki Cloud. This document tells you which models we use, what data we send them, what we don't do with your content, and how to turn AI off.

The AI features

• MindWiki AI chat — ask questions grounded in your vault. Retrieval-Augmented Generation (RAG) pulls the most relevant pages from your vault using semantic + keyword search, then feeds them as context to a frontier language model. The model's answer cites the pages it used so you can verify.
• Live voice conversations — talk to MindWiki AI hands-free. Speech-to-text transcribes what you say, the same RAG pipeline answers, and text-to-speech speaks the response back. A virtual avatar can lip-sync the response on iOS.
• Voice capture — dictate into your vault. Same speech-to-text engine, but the transcript is saved as a vault page (with optional audio attachment) instead of being routed to the AI for an answer.
• Vision (image analysis) — attach a photo to an AI chat. The model can describe what's in the image and relate it to your notes. Image analysis is currently rolling out — see "Current state" below.
• Autonomous agents — Auto-Linker proposes new wikilinks across your vault. Weekly Classifier reviews recent captures and proposes filing destinations. Pattern Detection surfaces cross-domain analogies. Monthly Auto-Report writes a vault-state summary. All proposals land in the Agents workspace for you to review and accept; agents never modify your vault without your approval.
• MindWiki AI Vault Master (macOS Power tier) — autonomous reorganization, refactoring, and filing operations with confirmation before write.

Which models we use

Where possible we route to frontier foundation models from Anthropic (Claude family) and OpenAI. Embeddings, lighter classification jobs, and most search ranking run on Cloudflare Workers AI. Voice is powered by Deepgram (speech-to-text), Cartesia (text-to-speech), LiveKit (real-time transport), and Anam (virtual avatars).

Power-tier users get "frontier routing": hard questions go to the strongest available model regardless of cost; everyday questions go to faster, cheaper models. You don't pick the model; we pick it per question, and we may switch providers when better models ship.

What data the models see

When you ask MindWiki AI a question, the system sends three things to the model:

• Your question text.
• The most relevant pages from your vault (retrieved via RAG — usually 5 to 20 pages, depending on the question).
• A short system prompt that tells the model how to behave and how to cite sources.

For voice conversations, an audio chunk is sent to Deepgram for transcription before any text reaches the LLM. The transcript and your vault context then follow the same RAG path. For vision, the attached image is sent alongside the question text.

For autonomous agents, the agent sees only the pages it's asked to process (the recent capture batch, your full vault index for Auto-Linker, etc.) plus a system prompt describing its task.

We don't train on your content

We do not use your vault content to train AI models. We do not share your vault content with model providers for training. Our contracts with Anthropic and OpenAI use the API tier — both providers have committed not to train on API data by default.

We don't sell or share your vault content with anyone for advertising. We don't maintain a content lake of user data. Every API call is stateless from the provider's perspective; nothing is retained beyond standard short-term inference logs which the providers use for abuse-prevention and which they delete on their own retention schedule.

Subprocessors

The AI providers that may receive snippets of your vault content during an AI query:

• Anthropic — LLM inference. Privacy policy.
• OpenAI — LLM inference (when routed). Privacy policy.
• Cloudflare — Workers AI for embeddings and lighter classification. Privacy policy.
• Deepgram — speech-to-text transcription (voice conversations and voice capture). Privacy policy.
• Cartesia — text-to-speech for AI voice responses. Privacy policy.
• LiveKit — real-time audio transport between your device and our voice agent. Privacy policy.
• Anam — virtual avatar video synthesis. Only the TTS audio stream is sent to Anam; your vault content is not. Privacy policy.

Our overall list of subprocessors (storage, payments, infra) is in the Privacy Policy.

Accuracy and your accountability

AI output is generated, not retrieved. Even with RAG citations, the model can summarize incorrectly, paraphrase a source in a way that changes its meaning, or hallucinate facts that aren't in your vault. We tune for accuracy and surface citations specifically so you can verify, but you remain responsible for what you do with AI output.

Don't treat AI output as professional advice — legal, medical, financial, or otherwise — even if you've asked the AI to act in that role. MindWiki AI is a thinking tool, not an authority.

Current state of features

MindWiki AI features are versioned and shipped continuously. As of the date above:

• AI chat, voice conversations, voice capture, autonomous agents, and the macOS Vault Master are live.
• Vision (image analysis) is partially shipped: you can attach an image to an AI chat on iOS today, but server-side pixel analysis is rolling out. Until full vision lands, the model receives the image attachment as context but answers primarily from text.
• Frontier model routing is live for Power-tier accounts.

We update this section when a feature changes state.

How to turn AI off

Every AI feature has a per-toggle off-switch in the app settings.

• iOS — Settings → MindWiki AI. Toggle "Voice conversations" or "Vision (image analysis)" off to hide those entry points entirely.
• macOS / web — Settings → AI. Disable individual agents or block AI access on a vault subtree.

When you turn an AI feature off, no AI request for that feature is made. Vault content is not sent to model providers for features you've disabled.

Prohibited AI uses

The Acceptable Use Policy applies to AI features specifically. Don't use MindWiki AI to generate content that would violate it — illegal material, harassment, fraud, attempts to extract training data of other users, or to jailbreak / circumvent provider safety policies. The provider may refuse such requests at the model layer; we may suspend accounts that try repeatedly.

Contact

Questions about AI behavior, model choices, or data flow? Email support@mindwiki.io.