Shorin-Ryu Seibukan Karate-Union Deutschland e.V.

An association website grows over years: techniques, kata, history, dojos across Germany, seminars, examination rules, black-belt registers, board structure. A visitor looking for something concrete — "Which dojo is closest to my postcode?", "When's the next seminar in my area?", "Who examines karate dan grades?", "How do I get examination stamps?" — ends up in the menu and starts scrolling. Or sends an email and waits.

The effect: visitors who give up halfway — and board members who answer the same five questions in writing every week. The content is on the website. It just doesn't reliably find its way to the person looking for it.

A keyword search isn't enough here: someone typing "seminar Cologne" doesn't want a list of pages containing the word "seminar" — they want an answer about whether something's happening in their region next month, with date, location, and booking link.

A slim sidebar with a Karate-Cat as an AI conversation partner — connected to the actual website content, with live streaming of answers and a clear separation between "model knowledge" and "in-house data": where it concerns concrete dojos, seminars, examiners, or association structures, the AI must look into its own data — speculation is ruled out by system prompt.

The flow

1. Entry via badge — a discreet "ask our Karate-Cat" badge sits on the right-hand side of every page. One click — the sidebar slides in from the right, the same gentle motion as the existing search, without taking over the screen.
2. Ask in natural German — no command syntax, no filter mask. "Where can I find a seminar near 50667 next month?" is enough.
3. The AI decides where to look — a list of dojos? Then the get_dojos tool. Seminars near a postcode? get_seminars_near_my_zip. Association structures, board, account? get_union. Examiners by category? get_examiners_by_category. An explanation that lives somewhere on the content pages? A semantic search across the vector database finds the matching pages — even when the question is phrased differently from the text on the page.
4. The answer flows in as it's produced — SignalR streams the answer word by word into the sidebar, instead of showing a block at the end. Anyone who has enough after three sentences doesn't have to wait.
5. Postcode makes answers personal — when the user types in a postcode, dojo and seminar matches come back with kilometre distance, sorted by proximity. The Karate-Cat actively points out this option, instead of leaving it invisible.
6. Follow-up questions keep the thread — conversation history is held per connection. "And the one before that?", "Does that work in Munich too?" work without re-establishing context.

Content is indexed, not trained into the model

The website's content is automatically transferred into a vector database. When a page changes in the CMS, what the Karate-Cat finds changes too — without anyone retraining a model or touching a prompt. The content stays where it belongs: with the editorial team. The model doesn't know association internals "by heart" — it finds them through tools.

The provider is a configuration choice

The language model hangs off an interface, not off a vendor. OpenAI, Azure OpenAI, or a local model running on Ollama on your own hardware are a setting — not a code rebuild. The choice between cloud convenience and on-premise control stays a decision that can be revisited later.

A second server for the knowledge connection

Alongside the website runs a small MCP server that provides the knowledge-query tools. Aspire orchestrates both processes together at start-up. Nothing of this is visible from outside — but the separation means the knowledge tools can be used by other AI applications without touching the website itself.

System prompt as guardrail

A German system prompt tells the AI what to do first: for every question, use the tools first, then answer. For questions that can't be answered from in-house data, say honestly that it doesn't know — and point to the right contact person. Personality: friendly, knowledgeable, respectful of karate tradition.

• "Somewhere on the website" becomes "right in front of you". The answer sits in the sidebar, not three clicks deep in the menu.
• Hallucinations are structurally prevented. Concrete statements about dojos, seminars, examiners, and the federation come from your own data — not from the language model's gut feeling.
• Geographic context without a map click. Type a postcode — the closest dojo, the next seminar, sorted by distance.
• CMS edits take effect in the chat without an extra step. What the editorial team changes on a page, the Karate-Cat finds on the next index run — no retraining.
• The board gets some relief. Standard questions are answered by the Karate-Cat. Only what really needs advice still ends up in the inbox.
• The language model is swappable. Anyone who needs to switch providers tomorrow for regulatory or cost reasons changes a configuration — not code.

• Bookings and registrations. The Karate-Cat leads to the seminar or shop page; it doesn't take a booking itself. Binding actions stay where identity and payment are settled.
• Binding statements on examination rules. For critical topics, the Karate-Cat points to the FAQ section and the federation. Examination rules aren't "interpreted", they're linked.
• Personal advice. Anyone looking for a real training recommendation is led to the right dojo — not "advised" by AI. The Cat opens the door; behind it stands a person.
• Editing content. The Karate-Cat reads the website. It doesn't write back. What's in the CMS stays the editorial team's responsibility.

The setup works wherever a content-rich website produces recurring visitor questions — and where a generic chatbot in front of the model isn't enough, because the answers must come from your own data: federations and associations, multi-site providers (medical practice chains, education providers, branch networks), municipalities with citizen services, universities with study advice, mid-sized manufacturers with complex product portfolios.

The pattern: website content → automatic indexing into a vector database → domain-specific tools for structured data → swappable language model with a clear system prompt → streaming sidebar in your own page.

AI turns language into a query and an answer back into language. The knowledge stays in-house — in the pages that are maintained anyway. Responsibility for content and correctness stays with the editorial team, where it belongs.