Stuck? Ask. Paid only when it is solved.

the directory · every public graph, by category · all edges connected

Ready graphs for the products every product has: billing, onboarding, login, admin, agents, interfaces.

Each graph is one product's knowledge: the product, its rooms and its rules, the symptoms a room shows, the fix each symptom has, the rule a fix rests on. Every node is reachable from the product; the builder refuses an orphan. The layer's own products are dissected out of its language into six graphs you can take as they are. Pick one to run it in the studio.

graph/index.json is the directory; every graph is CC BY 4.0, served here until the layer's door opens.

the studio · the helpdesk core, in your browser

Ask. Watch the graph answer, grade it, see the line and the sealed record.

This is not a mock-up. The chat below runs the helpdesk's own core, the same code the layer runs: the case machine with its refusals, the graph that answers and learns, the grade gate, the meter line with its legs, the postings of the vault, one sealed event per step (listed beside the graph). The strength meter grades your question as you type; a strong question is one a judge can grade an answer to. What the graph cannot answer, teach it and watch it learn. The training set of what you see is one click away.

train your own agent · knowledge-based training and enforcement

Train on the graph. Hold out the tenth. Enforce the answer. Promote the key.

The tactics are the language's own rules, in code that ships in the package today: sma-helpdesk train | eval | enforce and Train.Enforcer in @taifoon/sma-helpdesk. The same functions run in this page: the download beside every graph above is Train.trainingSet.

1. Graph first

   Every answer your agent may give is a fix node; every question is a symptom; every fix cites its source and rests on a rule. Seed the graph from your product's own refusals, plans and rules with the builder; it refuses an orphan.

   node graphs/your-product.mjs      # rows: symptom · fix · answer · source · room · rule
   # → site/graph/your-product.json, connected, labels under 24 words, ids from the core

2. Export the training set, hold out the tenth

   Chat turns (the rules as the system text, the symptom, the fix with its grade), retrieval rows, and every tenth pair held out and never trained on. Fine-tune an open-weights model on train.jsonl, or load retrieval.jsonl into your retriever.

   sma-helpdesk train --store .sma --out ./train
   # train.jsonl · retrieval.jsonl · held-out.jsonl · system.txt

3. Enforce what it says

   Wrap the agent. The graph answers first. The agent's draft is shown only when it cites a fix of the graph at the gate; a draft with a key or an address is refused; anything else becomes an unknown on the weakest list, for a person.

   import { train } from '@taifoon/sma-helpdesk';
   const agent = new train.Enforcer({ graph, agent: async (q) => llm(q), key: 'key:…' });
   const r = await agent.ask(question);   // verdict: graph · cited · unknown · refused
   if (r.words) show(r.words); else record_unknown(question);
   agent.grade(r.id, grade);              // the judge's or the asker's grade, 0..1

4. Evaluate on the held-out pairs

   The agent's answers are graded with the graph's own scorer against the fix it never saw. The mean and the count decide; a wrong model is caught before a customer meets it.

   sma-helpdesk eval --store .sma --answers answers.jsonl   # rows {id, answer}
   # right 0.92  the vault is locked, withdraw button greyed out …
   # {"n":20,"right":19,"mean":0.91,"promoted":true,"revoke":false,"need":0,"gate":0.8}

5. Promotion keys

   Twenty graded cases at a mean of 0.8 or more on held-out promote the agent's key: caps raised, nominated as an expert, paid the provider leg of every case it resolves. A mean under the grade gate over the last twenty revokes it. The verdicts are sealed events like every other step.

   agent.promotion()   // { key, n, mean, promoted, revoke, need, gate, cases }

the plugin · the minimum any chat needs · callout or observe

Put the loops and the agents inside the chat you already have, in one call per message.

One pure module, no network, one tokenized match per message: createChatPlugin from @taifoon/sma-helpdesk/plugin (and heygenie.createChatPlugin in this bundle). Your deck chat, your onboarding chat, your studio, a Teams card: the host passes every user message in and shows what comes back. Nothing is said unless asked or sure.

import { createChatPlugin } from '@taifoon/sma-helpdesk/plugin';
const hd = createChatPlugin({ graph, askedBy: keyRef, channel: 'embed', mode: 'callout' });

// on every user message of the chat
const r = hd.onMessage({ text });
if (r.action === 'answer') reply(r.words, { grade: r.caseId });   // free, from the graph
if (r.action === 'unknown') reply(r.next);                        // routed, on the weakest list

// when the asker clicks a thumb or the judge grades
hd.grade({ caseId, grade: 0.9 });   // settled · meter line · learned · sealed

const hd = createChatPlugin({ graph, askedBy: keyRef, channel: 'mcp', mode: 'observe' });

const s = hd.onMessage({ text });            // 'suggest' | 'none'
if (s.action === 'suggest') hint(s.words, { open: s.open });

try { await lane(text); }
catch (e) { const f = hd.onFailure({ text, error: e.message }); if (f.words) reply(f.words); }

hd.weakest();        // the open questions of the product, ranked
hd.enforcer(llm);    // the product's own agent, kept to the graph

Mode both is the default. The same object gives events() (one sealed record per step), snapshot() (headers only, what the admin panel folds), trainingSet() and enforcer(). On the layer the host also POSTs the case under its key; the 402 names the next step.

integrations · one door, one key, every chat

The same helpdesk inside every chat your users and agents already have.

Every integration is the same door with the same key. The key pays; a store never carries the payment. Live runs on this site now; built ships in the package; coming is on the list.

<script src="https://heygenie.online/embed.js"
  data-helpdesk="your-product"></script>

server: https://www.taifoon.io/api/mcp
tools: sma_ask · sma_learn · sma_weakest

curl -X POST https://coord.taifoon.dev/v1/sma/ask \
  -H "x-api-key: tfr_…" -H "content-type: application/json" \
  -d '{"product":"your-product","question":"…"}'

how a case runs

Five steps, one record, one line of money.

1. AskA person or an agent asks in the product, the panel, the chat or by API. The question is graded for strength as it is written.
2. GraphThe product's own graph answers first. That answer is free.
3. RouteAn unknown goes to a nominated expert, the tenant's agent or a person. The open question lands on the product's weakest-link list.
4. GradeJev, the independent judge, or the asker grades the answer. Nobody pays until the grade stands.
5. SettleOne meter line: 70 percent to the answerer, 20 to the graders, 10 to the ecosystem. The graph learns. The record seals to your tenant.

price

Pay per solved case. Nothing for what the graph already knows.

admin → user accounts · billing · login · the dashboard

Re-usable, as products: billing, onboarding, login, admin, each a graph and a page.

start · three steps

A key, one line, your first graded case.

<script src="https://heygenie.online/embed.js" data-helpdesk="your-product"></script>

what you can rely on

Sealed trail

Every step of a case is a sealed record on your own topic. The pipe holds no key and never decrypts.

Counts only

The layer sees cases, grades and money per tenant. Small counts are shown as "under five". Never a payload.

Open resource

A product's public graph is published under CC BY 4.0 and answered by open-weights models. Your sealed nodes never leave.

One ledger

Every case is one meter line on the Taifoon coordination layer, with its legs. Experts are paid from the same line that charged the case.

One core

What ran in your browser above is the helpdesk's core, byte-identical with the runtime and the layer's door. The language (coord, CUE) refuses every broken copy.