AI course drafts & resource answers

Create with your expertise

1. Open Course Studio and choose Draft with AI. Your workspace must be explicitly enabled by the platform operator.
2. Describe the learning goal and audience. Include only information you are allowed to send to the external model provider; avoid personal and confidential data.
3. Generate a proposal. The assistant creates 2–3 modules with text lessons and multiple-choice knowledge checks.
4. Expand every lesson and review the answers, accuracy, suitability and rights. Acknowledge the review and choose Open as editable draft.
5. Refine the accepted content in Course Studio. Normal draft autosave applies; preview and publish separately when ready.

Ask your resources

Quotations are checked against selected source text. This does not prove the answer’s interpretation is correct. Linked resources may change later; historical source snapshots and conversation history are not retained.

This pilot uses existing plain-text resource entries. PDF/DOCX ingestion, OCR, whole-library search, learner AI tutoring and autonomous actions are not included.

1. Open Resources and choose Ask your resources. The operator must enable AI_RESOURCE_ASSISTANCE_ENABLED for this author pilot.
2. Choose one to three current text resources, ask a specific question and confirm that you are authorised to send the selected text to the AI provider. The combined limit is 32,000 UTF-8 bytes.
3. Choose Find an answer. Read each quotation and follow its source link. If the selected text cannot answer the question, the assistant abstains.
4. For a source-based course, choose the same resources in Draft with AI. Review the proposal; source versions are checked again before it opens in the editor. Refresh and regenerate if a resource changed.

Your content and control

Your brief/audience or question and any explicitly selected resource titles/text are sent to the explicitly configured AI provider. No other resources, existing courses or learner records are attached. Selecting sources requires permission for external processing and explicit consent. Provider privacy restrictions are requested, but this is external processing subject to provider and account policies.

Generation never saves or publishes a course. The proposal remains in browser memory until you accept it. Reloading or leaving the page loses an unaccepted proposal. Accepted drafts use normal course storage and versioning.

AI can make mistakes. Structured output checks format and limits; it cannot guarantee factual accuracy, correct answers, legal compliance or educational quality.

Setup and limits

• Operators configure AI_DRAFTS_ENABLED, AI_ALLOWED_ORGANISATIONS (exact company UUIDs), AI_MODEL and AI_API_KEY_FILE on the API server. API keys never belong in frontend variables.
• Only active organisation-wide authors, administrators and owners can generate. A team manager or learner needs an author grant.
• The brief allows 20–4,000 characters; audience allows 2–200. Each workspace defaults to 20 shared attempts per UTC day, subject to configured company and installation limits, including failed and cancelled requests. Requests can take up to 75 seconds.
• There are no automatic retries. Cancelling may still incur provider usage. An interrupted request cannot be recovered; a new request consumes another attempt.
• A provider outage, unsupported model, unavailable privacy-compatible endpoint or invalid output produces an error without changing existing courses.

API and roadmap

The API reference lists GET /api/v1/ai/drafting and POST /api/v1/ai/course-proposals under AI assistance. These endpoints use a browser session; generation also requires Origin and CSRF validation. The server determines the company from the session.

The local author pilot includes course drafts, selected-resource answers and a source-to-draft review workflow. The API also provides POST /api/v1/ai/resource-answers and POST /api/v1/ai/sources/verify. Answers and course requests share the daily quota. LangChain and LangGraph are not installed: these bounded Go flows do not yet need durable graph orchestration. AI tutors, general document ingestion and autonomous workflows remain separate future work.