AI providers and data flow
What each Vidext feature may send to an AI, media, or research provider, and which providers Vidext uses.
This page tells you what leaves Vidext when a feature uses AI: which data each feature may send, to which providers, and what Vidext keeps afterwards.
Vidext uses different providers for language, document reading, video understanding, transcription, speech, images, research, and presenter video. A request goes only to the service the feature needs; it is not broadcast to every provider.
Provider reference
This page is Vidext's reference for the AI, media, and research providers it uses. It describes data by feature, without tying the documentation to a model version that may change.
Vidext's commitments
- No Vidext training on Customer Content. Vidext does not use Customer Content to train its own AI models. The applicable agreement and Data Processing Agreement govern provider use of that content.
- Customer ownership. You retain your rights to the material you provide and the content generated for your organization, subject to applicable law.
- Purpose-limited processing. Content is processed to perform the requested feature, provide support, maintain security, and meet legal obligations.
- No sale or rental. Vidext does not sell or rent Customer Content or personal information.
See the Terms of Service and Data Processing Agreement for the binding language.
How a request moves
- You provide or authorize content. This may be a message, a document, an image, an audio or video file, a video link, your voice while you dictate, or a website.
- Vidext selects context. The feature combines your instruction with the source material, conversation history, and system instructions needed for that task. In the chat, Vidext's agents read sources as they need them instead of receiving all of them at once.
- Routing and safety controls run. The request is sent to a suitable provider. A configured fallback may handle the same task if the primary provider is unavailable.
- Vidext stores the result. Outputs and the operational records needed for history, editing, delivery, usage accounting, reliability, and security are saved by Vidext as described in Data handling.
Vidext does not automatically send the whole organization or knowledge base with each call.
What each feature may send
"May send" lists the data categories a feature needs; a single request rarely contains all of them.
| Feature | What may be sent |
|---|---|
| Building in the chat | The chat is where admins and members tell Vidext's agents, the AI agents that plan and write courses, what to teach. What is sent is the conversation: your messages and attachments, the passages of sources and research the agents read, the course plan and scripts, and what the agents remember: your lasting preferences and the facts about the organization you confirmed. It is sent each time the agents respond, and when the agents tidy that memory, it is sent together with the organization's files to the model that tidies it. A message that tags a colleague and carries no file or skill is a note between people: it is not sent. |
| Reading documents | Uploaded documents and page images, to extract their text; attached photos, to check whether one can be a presenter; the extracted text, to build the search index. |
| Video and audio sources | Uploaded videos and video links, to describe what they show and to transcribe them with speakers told apart; uploaded audio, to transcribe it. |
| Dictation | Your microphone audio while dictation is on in the chat. It streams straight from your browser to the speech provider with a single-use credential, and does not pass through Vidext. |
| Research | Search queries, public web addresses, and research keywords. Workspace documents are not sent. |
| Your company website | Public pages of the website, to build the brand kit and the company profile. |
| Narration and voices | The script text to be spoken and the chosen voice and settings; a short sample text when you preview a voice. |
| Images | Image instructions and reference images, which can come from your sources. |
| Presenter video | The presenter image, which can be a photo you sent in the chat to make your own presenter, and the narration audio for that video. Only when you choose portraits for a new presenter, that photo, the one-sentence setting the agent writes for each portrait and up to nine of Vidext's own presenter stills, sent to the image model to draw three portraits; the portraits you do not pick stay stored with the organization's presenters and are shown nowhere. |
| Course languages | The course content being translated. |
| AI tutor (learners) | The learner's message, the conversation, and the course content relevant to it; a web search request when the learner asks about something beyond the course. |
Depending on the feature, these requests go to OpenAI, Google, Anthropic, xAI, Mistral AI, ElevenLabs, Inworld AI, RunPod, OpenRouter, and Cloudflare for models and media, and to Firecrawl, Exa, and OpenAlex for research.
When you build in the chat, the agents may write some of their work, such as module plans and module scripts, with Anthropic's Claude Haiku, reached through OpenRouter. Those requests may carry what is listed for building in the chat above. They leave the conversation's isolated environment through Vidext's Cloudflare AI Gateway, with payload collection disabled, and OpenRouter passes them to a provider that serves that model.
Documentation assistant
OpenRouter also routes the model used by Ask AI on this documentation website. It receives the documentation-chat history, the current page URL, and excerpts returned from the public documentation search index. That assistant is separate from the Vidext product and has no automatic access to an organization's workspace, sources, conversations, or learner data unless a visitor manually pastes that information into the documentation chat.
Routing and observability
Cloudflare AI Gateway routes model and media requests and returns provider metadata. Vidext sets payload collection to disabled on these requests. Provider, model, token or character usage, cost, status, duration, and trace identifiers may still be recorded.
AI Gateway does not replace the downstream provider: a routed request still reaches the model or media provider named above.
When a request to OpenAI is turned away for exceeding a rate limit, such as the capacity Cloudflare gives the gateway, the same request may go through OpenRouter instead. OpenRouter passes it to the same OpenAI model, and Vidext allows OpenAI as its only upstream provider for these requests. OpenRouter then receives the data the feature would have sent to OpenAI.
What Vidext stores
Vidext keeps the results and the records it needs to run the product: conversations, sources and what processing made from them, generated media, and operational records such as provider, model, usage, cost, status, and trace identifiers. Data handling lists each category.
Retention and deletion
Vidext storage and temporary provider processing are separate.
- Content stored by Vidext follows the applicable account, contract, and legal retention rules described in the Privacy Policy.
- Providers may temporarily process or retain request data for delivery, safety, abuse prevention, or legal compliance under their applicable terms. The provider links above lead to the published policies that describe how each service handles data.
- Deleting an item in Vidext does not imply that every temporary provider copy or backup disappears at the same instant. Requests for access, correction, export, or deletion can be sent to datos@vidext.io.
Sources of truth
Last updated on