Technical overview
How the platform is designed to work, from the manual you have today to content that people and AI assistants can use.
In development The platform is in development. This page describes its design.
01 · Principles
Each task, concept, error code or question for a specific model gets its own stable web page.
The first two or three sentences answer the question completely and name the brand and the model.
Each topic exists once as structured content. The web page, Markdown, video, subtitles, audio and MCP response are made from it, so they do not contradict each other.
Every topic has one canonical address, on your own domain or on your subdomain of the platform domain.
Assistants read text. Video and audio help people, and count for assistants through the text around them: titles, transcripts, subtitles and chapters.
02 · Content model
Topic types: task, concept, reference, troubleshooting, error code, FAQ. Each topic version is immutable and carries a SHA-256 source hash.
03 · Import
| Source | How | Result |
|---|---|---|
| DITA/XML | Mapped field by field: steps, supplies, warnings, map order, links | Topics |
| Text and layout extraction, then segmentation into topics | Drafts to review | |
| iiRDS | Packages read as untrusted input, mapped per client | Topics |
| Word, Markdown | Planned; the exact list of upload formats is still open | Drafts to review |
Each topic keeps an import key (source, source id, language). A new version never duplicates; only changed topics and what depends on them are rebuilt.
LLM help with structuring runs through your approved provider, produces drafts only, and never rewrites safety text.
04 · Review workflow
| Role | Does |
|---|---|
| client-admin | Domains, SSO, crawler policy, public marking, MCP, YouTube connection |
| author | Imports, edits drafts, submits for review |
| reviewer | Approves or rejects, never their own edits |
| publisher | Publishes and retires, as a separate step after approval |
| viewer | Reads everything, sees status and analytics |
Every transition is written to the audit log. Warnings, cautions, torque values and part numbers are never reworded without human approval.
05 · Generate
Web
Search
Media
Idempotent. Each job is keyed by topic version, format and host, and skips work when the source hash is unchanged.
Text never waits. The page goes live first; video and audio attach when ready.
Stale is detected. A changed step marks only its video, audio and subtitles for rebuild.
06 · The topic page
Complete without JavaScript.
07 · Where content lives
https://{host}/{model-group}/{lang}/{topic-slug}…/{topic-slug}.md Markdown, noindex, canonical link…/{error-segment}/{code} error-code topicshttps://{host}/media/{model-group}/{lang}/{slug}.mp4Your own domain, or {slug}.ifuto.ai delegated by DNS to your own deployment. That host is always canonical.
A slug change writes a 301. A retired topic returns 410 with a link to its replacement. URLs are never deleted.
robots.txt, sitemaps with hreflang and video, llms.txt, llms-full.txt, and mcp.json when MCP is on.
08 · Honest about each channel
Assistants query official content directly and get the current version plus the URL to cite, without waiting for a search index.
Strong reach for how-to questions, and physical steps are clearer on screen. Only when it matches its page one to one: title, subtitles, chapters, link back.
Rarely found or cited by assistants on its own. On the page it helps accessibility and hands-free use.
Major crawlers mostly read the HTML. Agents and developer tools do use it, so it is provided.
The web page underneath all of them is what assistants read.
09 · MCP server
| Tool | Returns |
|---|---|
| search_topics | Ranked topics for a query, by brand, model, language or type |
| get_topic | Steps, warnings, supplies, video chapters, Markdown |
| lookup_error_code | Meaning, cause and fix for a code on a model |
| list_models | Models with GTINs, aliases, languages, topic counts |
| resolve_product | Best match for free text or a GTIN, with confidence |
Every result carries the canonical URL, update date, valid models and publisher.
Ambiguous model: candidates, never a guess. Wrong model: no steps, a pointer instead.
Safety text returned exactly as approved.
Public and read-only, per client deployment, off until you enable it. A second, authenticated endpoint behind your SSO serves drafts and authoring tools, under the same four-eyes rules.
10 · Media pipeline
Built from the topic: summary plus steps, a chapter per step group
Text-to-speech by your chosen provider, one voice per language
From word timings; chapters start at the step anchors
Title card, a clip per step, on-screen labels and warnings, end card
Your own channel, captions, a playlist per model group
Per-step clips meet. The last frame of one step's clip matches the first frame of the next; every seam is checked.
Old versions go. When narration changes, the new video goes up and the old one is deleted on every channel.
11 · Measure
Which AI and search bots fetch which topics, and which topics are never fetched. Built from edge logs of public hosts only, counted per bot family.
Off until you enable it. No IP addresses kept; raw logs deleted within 24 hours.
A curated set of real user questions, run regularly against assistants with web search. Records whether the answer cites your page and whether key facts match.
Opt-in, from your own deployment and account. Results are trends, not guarantees.
Next step