Cortex Trainings is a standalone application that turns the knowledge already in a Cortex instance into interactive courses. It runs as its own deployment and connects to the backend exclusively through the REST API with a read-only key — a standalone app in the same sense as Cortex Chat, as opposed to Apps hosted inside an instance.
The pitch: describe a topic and an audience, and an agent researches your knowledge base and writes a complete curriculum — levels, learning objectives, scripts, quizzes, and every source it used, cited with dates. You review and approve that as a plain document. Only then does production start, generating cinematic AI video, a narrated voiceover in your language, beat-synced explainer animations and gamified interactions, and delivering the finished course as a single HTML file that runs offline on double-click. Onboarding, compliance and product training built from what your company already knows, in about half an hour.
The two-part workflow
The split between writing and producing is the core design decision, because a text change before production is free and the same change afterwards means a new voiceover, a new render, and possibly a new film.
Part 1 — Curriculum (cheap)
- Four required inputs — topic and learning objectives, audience and prior knowledge, content language, and desired duration. The app refuses to guess any of them; duration determines the number of levels (3–4, 5–6, or 7–8).
- Optional inputs — existing material to build on, a collection to scope research to, and up to three character plus three style reference images, analysed once on upload and binding on every prompt from the curriculum onward.
- Research-first — the agent issues a broad fan-out of deep-research queries against your instance before writing, then further queries to fill gaps as it drafts. Knowledge-base queries are cheap relative to agent tokens, so the design spends generously here.
- A reviewable document —
curriculum.mdcontains a fact sheet, learning objectives, a level table, and per level the teaching text, voiceover script, media plan and fully written-out interaction, plus a final check, a cheat sheet, cited sources and a production estimate. - Free revisions — request changes in a chat; each revision saves a new version and propagates through the objectives, level table, final check and sources.
- Honest gaps — when the knowledge base doesn't cover something, the curriculum names the gap and asks, rather than filling it with plausible general knowledge.
Part 2 — Production (costs money)
- Approval-gated — unreachable until you explicitly approve. From that point the curriculum is binding: what gets produced is what the document says.
- Resumable pipeline — eight steps with live progress over SSE, ending in an automated browser click-through that fails the run if an interaction cannot be completed. Completed steps are skipped on resume and every artefact is cached, so a failure never re-pays for finished work.
- Two human checkpoints — you choose the guide character's reference image (the anchor the storyboard's character shots are generated from), and you confirm a live cost quote before any video is generated.
- Media — generated video for story moments, locally rendered animations for concepts and lists, and images for interaction screens. Film shots are de-padded before assembly, because video models compose to their own taste and pad to the requested ratio.
What a training unit is
A single self-contained HTML file, typically 10–15 MB, with all media embedded. No server, no build, no accounts, no telemetry — it runs on double-click and is shareable by mail, Drive, or an LMS.
- Levels — each pairs one medium with one interaction; the interaction must be completed to advance, videos are skippable.
- Interactions — multiple choice, myth-or-fact cards, sorting tasks, find-the-mistakes, self-assessment sliders, and a shuffled final check. A different form per level, because repetition is what makes e-learning a chore.
- Gamification — XP with reduced points on a second attempt, level badges, a progress bar.
- Progress — saved in
localStorage, so closing the tab and returning resumes where the learner left off. - A printable cheat sheet — the summary screen carries every key takeaway with print styling.
- Any language — learner-facing text and the voice follow the training's language; the
langattribute is set accordingly.
A guide character anchors the visual identity. By default it is an abstract object — a glowing orb, a crystal, a cube — never a human, because abstract objects stay consistent across image generations where a person drifts between shots; its colour is then the configured accent, applied throughout.
Or bring your own. Upload up to three reference images of your own character — a mascot, a figure, a product — and a vision model extracts a precise description of it, which is then used to generate the anchor image from your actual images rather than from a paraphrase. Your character keeps its own colours: the accent still rules the interface, animations and highlights, but nothing is recoloured to match it. The consistency caveat above still holds, so a detailed human character drifts between shots more than a stylised or abstract one.
Upload up to three style reference images too, and their extracted aesthetic replaces the preset visual style in every film and image prompt. With both supplied, your character is rendered into the world the script describes, in the look your references define.
The guide does not appear in every shot, and that is deliberate. The storyboard decides per shot: roughly half a film's shots — establishing shots, object details, concept imagery — are character-free, and the opening shot almost always is. A guide in every shot reads as a mascot parade and buries the subject the training is actually about. Character-free shots still carry the shared style, so they stay in the same visual world.
Choosing the medium matters
Video is priced per second and dominates the cost of a training by an order of magnitude; animations are rendered locally and cost nothing. Animations also render razor-sharp text in any language, which generative video cannot do — so concepts, lists, rules, processes and numbers belong in an animation, and film is reserved for the two or three moments that genuinely need emotion. A curriculum will often suggest this trade-off itself.
Access and configuration
Cortex Trainings only ever reads your instance.
- Read-only key — use a
cortex_ro_…key, ideally scoped to the collections a training may draw on, so a leaked key exposes only content you already chose to expose. - Required instance flags —
ENABLE_AGENTIC_RAGandENABLE_AGENT_RESEARCH, because the research phase relies on agentic retrieval. - Endpoints used —
POST /api/search,POST /api/ask/stream, the community and graph endpoints for mapping a domain, andGET /api/documents/{id}/contentfor reading primary sources. - Server-side only — the key never reaches a browser, so the app's origin does not need to be on your instance's CORS allowlist.
Configuration is entirely environment variables: instance URL and key, the AI provider key and model choices — agent, vision, image, image-editing, video and speech — interface language, accent colour, and a storage path. Generated projects are stored as plain files — every curriculum version, the production plan, per-step state and all media — so runs stay inspectable, resumable and recoverable.
Requirements: Node 22+, ffmpeg/ffprobe, and a headless Chromium for rendering animations.
Where to go next
- Apps — the other kind of app: small web apps that run inside your instance
- Cortex Chat — the other standalone app, for everyday question answering
- Ask AI — the retrieval and deep-research capabilities the curriculum phase builds on
- Authentication — API keys, permissions and collection scoping